1 Introduction

The combination of AI tools with neurotechnology has opened the door to a large range of novel applications. Recent innovations include brain-computer interfaces (BCIs) that enable patients with amyotrophic lateral sclerosis to compose Tweets solely with their thoughts, and aphasic stroke patients to talk live via neuroprosthesis at a speed of 83 words per minutes [1].

In recent years, there has been an increasing number of news reports about purported AI applications that putatively decode brain activity for mind-reading purposes. These putative AI mind-reading capacities have been portrayed in news media as spectacular yet concerning advances. Recent media headlines such as “The brain is the final frontier of our privacy, and AI is about to breach it (Yahoo News)” [2], “Mind-reading technologies have arrived (VOX)” [3], “AI makes non-invasive mind-reading possible by turning thoughts into text (The Guardian)” [4], “This ‘mind-reading’ AI system can recreate what your brain is seeing (Euronews)” [5], “AI-Powered ‘Thought Decoders’ Won’t Just Read Your Mind—They’ll Change It (Wired)” [6] are so frequently portrayed that one feels ‘AI ability to read the mind’ is mainstream reality. However, given that news media often portray BCIs with overall positive and sensationalist enthusiasm [7, 8], one could be reluctant to subscribe to the belief that AI can access and decrypt hidden aspects of the mind.

Nevertheless, these claims about the putative AI mind-reading ability are also echoed within the world’s most respectable and influential institutions. For example, The International Bioethics Committee of United Nations Educational, Scientific and Cultural Organization (UNESCO)’s report on “The Risks and Challenges of Neurotechnologies for Human Rights [9]” highlights the multifaceted impacts of combining AI and neurotechnologies capable of “reading” and “writing” brain activity. Further, even in the most impactful academic journals, narratives about the putative AI ability to read the mind are conveyed in titles such as “Artificial intelligence is learning to read your mind—and display what it sees (Science)” [10]; ““Mind-reading devices are revealing the brain’s secrets” (Nature) [11]”. This growing body of claims appear to indicate that brain decoding may lead to mind-reading [1, 12].

For many academics, the increasing ability to decode brain activity—and potentially its associated thoughts—raises profound and unparalleled ethical concerns. This includes assertions such as: neurotechnologies allow “access, via a mind-reading device, to citizens’ specific thoughts” [13] or “companies and governments are developing devices that would allow people to communicate by thinking, to decipher others’ thoughts by reading their brain data, and to have access to all of the internet’s databases and capabilities inside their minds” [14]. These alarming ethical assertions follow a logic: If an individual’s neural data can be read or interpreted with or without their explicit consent, then this could potentially lead to degrees of surveillance and control over an individual’s innermost thoughts and convictions. Therefore, these scholars argue that we need new human rights —neurorights— to protect mental privacy, free will, agency, consent, etc. [13,14,15,16,17,18].

It is difficult to dispute that the integration of AI with neurotechnology has catalysed the rapid and unprecedented development of BCI discoveries and applications, introducing a range ethical and legal questions: Is there a right not to have your mind read? Do we have the right to control who accesses our thoughts? Should we consider establishing neuro-specific human rights to safeguard our mental privacy and integrity? Are we truly entering an epoch where our mind can be read by computer?

Central to these concerns is the fundamental question of whether the current state of BCI progress is truly reflected by the narratives portraying the putative AI ability to read the mind. Are those claims hyped by the booming field of AI reporting? If there is hype, should we conclude that conceptual works alluding the potential to read minds, such as new human rights, should be dismissed [19, 20]? To answer this fundamental question, it is crucial to critically examine the scientific evidence and academic literature about the supposed mind-reading capabilities of BCIs.

For that purpose, we conducted an interpretive content analysis, via a thematic framework [21, 22] of the academic literature to investigate and identify prevalent claims, trends, and technologies associated with the putative AI ability to read the mind.

2 Methods

We utilised both qualitative and quantitative methods within an interpretive content analysis framework [21, 22] to collate research trends and identify literature narratives in the field of brain-reading and mind-reading technologies. Interpretive content analysis is a method for quantifying interpretations of latent content within a context, e.g. academic corpus. In this method, researchers go beyond counting the frequency of the most straightforward denotative elements to make replicable and valid inferences from texts [21, 22]. This method provides a comprehensive understanding of the academic narratives, going beyond what is explicitly stated in the articles to examine the ways language is employed to construct meaning or perpetuate certain beliefs [21, 22].

Interpretive content analysis is an effective tool for making sense of a large corpus of academic literature. In our study, this method allowed us to examine the contexts in which the terms “mind-reading” and “brain-reading” are used within academic literature, as well as the meanings that can be inferred from their use. Interpretive content analysis facilitates a critical examination of the language and ideas surrounding AI and neurotechnology, revealing underlying patterns and assumptions within the field. It is particularly suited to exploring how neurotechnologies are represented in academic discussions and to uncover any claims or assertions about their potential impacts and applications. Our approach involves an interpretive content analysis aimed at characterizing the formation of beliefs in the academic literature, both in terms of its content and the range of its concepts [23].

We used simple, specific search terms to ensure precision and relevance in capturing the narrative directly related to these concepts. The academic corpus on which our discourse analysis is based includes studies that explore wide-ranging research questions and encompass various study designs (e.g. first-hand empirical studies, review studies, and conceptual studies).

2.1 Search protocol

We performed a Google Scholar search to constitute our academic corpus by identifying articles making claims about mind-reading through brain-reading. Google Scholar is a freely accessible web search engine provided by Google that indexes the full text or metadata of academic literature across various disciplines. Google Scholar offers a broad range of scholarly literature, reflecting the diversity of academic discourse and facilitating interpretive content analysis. Its vast resources provide rich data for identifying and interpreting patterns within qualitative research.

We used the Boolean search string “mind reading” AND “brain reading”. This returned articles from first occurrence until November 2023. In order to be included in our analysis, articles yielded by this search had to be written in English. For this content analysis, we also excluded articles as irrelevant if they were: identified as duplicates; inaccessible due to broken links or empty files; published in non-peer-reviewed formats such as personal blogs or US patents; or about a different field of research not concerned with mind-reading through brain-reading.

2.2 Analysis protocol

To encode the relevant articles, we performed an in-depth assessment of their features and recorded the results along with their titles, sources and dates of publication. Coding proceeded in two stages according to the analysis protocol in Fig. 1.

Fig. 1
figure 1

Interpretive Content Analysis Protocol and Stage Breakdown

Full size image

In the first stage, a researcher (IR) analysed each article for keywords and recorded the answers to a series of set questions in a spreadsheet using a binary coding system: a “1” for ‘yes, and a “0” for “no” (Table 1). A second researcher (FG) cross-checked coding at different intervals for validity. The set questions pertained to five different aspects of the articles’ claims: (a) claims about mind-reading, (b) claims about brain-writing, (c) ethical considerations, (d) type of neurotechnologies, and e) acknowledgement of hype.

Table 1 Interpretive Content Stage 1 Analysis Questions
Full size table

In the second stage, a researcher (IR) further analysed those articles suggesting that mind-reading is already possible to classify what was meant by ‘mind-reading’.

3 Results

3.1 Summary

Our search terms yielded 1017 articles published between 1972 and 2023 (November), of which 448 were deemed to be irrelevant based on our coding criteria methodology. Consequently, our interpretive content analysis and results are derived from the remaining 569 articles that met our relevancy standards. The set of relevant articles were published between 1999 and 2023 (November).

We note a progressive increase in academic discussions surrounding brain and mind-reading, as demonstrated by the number of articles connecting brain- and mind-reading in Fig. 2.

Fig. 2
figure 2

The Increase in the Number of Articles Connecting Brain-Reading and Mind-Reading by Year

Full size image

3.2 Stage 1A—Mind-Reading claims

Out of the entire collection of articles deemed relevant to our study, we discovered that 46% assert or imply the current feasibility of mind-reading, while a closely comparable 45% propose that mind-reading is not achievable at present but may become viable in the future, as detailed in Fig. 3. Analysing the trend up to 2015, there was a clear predominance of articles claiming the immediate possibility of mind-reading over those suggesting its future potential. However, this trend has recently shifted (Fig. 4). Notable spikes in the number of publications can be traced back to groundbreaking research papers, such as Kamitani and Tong’s [24] work in 2005, which garnered 2091 citations, and Norman et al.’s [25] review in 2006, receiving 2666 citations. This trend was further influenced by publications from Haynes in 2011 [26] and 2012 [27] and gained significant attention, likely boosted by sensationalist claims (Table 2). The highest cited articles at this time were from Yarkoni et al. [28] (2996 cites) and Naselaris et al. [29] (792 cites). Neither claimed mind-reading is possible but allowed for it to be possible in the future.

Fig. 3
figure 3

Claims About Mind-Reading in the Academic Literature

Full size image
Fig. 4
figure 4

Mind-Reading Claims in the Academic Literature by Year

Full size image
Table 2 Quotes Illustrating Claims that Mind-Reading is Possible Now
Full size table

Table 2 showcases an extensive collection of illustrative examples and verbatim quotations taken and systematically ranked in order of citation frequency from the 46% of Fig. 3. Each of these excerpts supports and substantiates statements and claims proposing the feasibility of mind-reading through the application of advanced, cutting-edge technological methods.

Table 3 offers a detailed compilation of select examples and direct quotations taken from the 45% articles in Fig. 3, ranked in order of the most cited. These excerpts specifically affirm and support statements and claims suggesting that mind-reading might be possible in the future, illustrating the range of perspectives and evidence presented in the academic literature on this topic.

Table 3 Quotes Illustrating Claims that Mind-Reading Might be Possible in the Future
Full size table

3.2.1 Terminology: inconsistent lexicon

As suggested through the data presented in Table 2 and Table 3, there appears to be a lack of consensus in the academic literature regarding the putative ability of BCI to read the mind, precisely what the technologies do (verbal lexicon) and what they are accessing in the mind (mental lexicon). Table 4 further elaborates by listing common keywords and phrases, which are employed in various combinations across different articles to describe the functionalities and capabilities of neurotechnologies. In our analysis of these articles, it becomes evident that the specific claims made by authors are not always explicitly clear. For instance, while some authors might use the phrase ‘read the brain’ in the context of identifying hidden intentions, others might use ‘access the inner sanctum’ to imply the detection of visual percepts, demonstrating the varied interpretations and applications of these terms in the field.

Table 4 The Verbal and Mental Lexicons in the Academic Literature
Full size table

3.2.2 Stage 1B—Brain-writing claims

Expanding upon previously gathered data and spurred by UNESCO’s reference to 'brain-writing' in their report, we conducted a focused analysis of the potential usage of this term in academic papers. Our observations reveal a notable and steep increase in the use of ‘brain-writing’ in conjunction with mentions of BCI and AI, particularly marked from the year 2019 onwards.

Out of the 569 articles we identified as relevant, Fig. 5 vividly illustrates a significant surge in the frequency of mentions relating to AI and BCIs within the context of mind-reading articles.

Fig. 5
figure 5

The Rise of ‘Brain-Writing’ Against BCI and AI Mentions in the Academic Literature

Full size image

3.2.3 Stage 1C—ethical considerations

Out of the 569 articles deemed relevant to our study, various ethical considerations were highlighted, with each being mentioned by a specific percentage of these articles (Table 5). It is important to note that some articles discussed multiple ethical considerations, contributing to the diversity of ethical themes observed in this body of literature.

Table 5 Ethical Considerations Raised in the Context of Mind-Reading Claims
Full size table

Figure 6 shows the growing ethical considerations raised in the context of mind-reading.

Fig. 6
figure 6

Growing Ethical Considerations in the Context of Mind-Reading

Full size image

3.2.4 Stage 1D—neurotechnologies

Table 6 demonstrates a range of technologies, each mentioned by a specific percentage of our set of relevant articles. It is noteworthy that some articles referred to more than one type of technology, indicating a variety of technological focuses across these studies.

Table 6 Neurotechnologies Mentioned in the Context of Mind-Reading Claims
Full size table

Figure 7 shows how the focus on particular neurotechnologies has changed in recent years. The number of mentions of fMRI appears to have peaked, while the number of mentions of BCI, ANN and AI is increasing.

Fig. 7
figure 7

Changing Focus on Neurotechnologies in the Context of Mind-Reading Claims

Full size image

3.2.5 Stage 1E—hype

Of the 569 relevant articles, 121 (21%) suggested a degree of hype regarding mind-reading claims. When tracked over time, it appears that the number of articles suggesting hype is increasing (Fig. 8). Figure 9 shows the frequency with which the 121 articles acknowledging hype mentioned each of the tracked neurotechnologies.

Fig. 8
figure 8

Acknowledgement of Mind-Reading Hype in the Academic Literature

Full size image
Fig. 9
figure 9

Breakdown of the Neurotechnologies Mentioned in Articles Acknowledging Hype

Full size image

Given these data and the changes in the frequency with which each technology has been mentioned over time in the academic literature (Fig. 7) it is possible to plot each technology against Gartner’s Hype Cycle (Footnote 1).

fMRI is a key technology in most hype claims (75%), yet its prevalence in the set of relevant articles appears to be decreasing. For this reason, it may be conceived as being past the peak of inflated expectations. BCI, AI and ANN are referenced less frequently in hype claims, yet their prevalence in the set of relevant articles is increasing. For this reason, they may be conceived of as on the upward slope of inflated expectations (Fig. 10).

Fig. 10
figure 10

The Hype Cycle of Mind-Reading Neurotechnology Claims

Full size image

BCI, AI and fMRI are now mentioned with a similar degree of frequency in the academic literature Fig. 7.

3.2.6 Stage 2: what is meant by mind-reading?

Further investigation of the 264 (46%) of articles suggesting that mind-reading is already possible revealed what was meant by ‘mind-reading’. The 264 articles were analysed for the types of mind-reading claims, using the categories listed in Table 7 and depicted in Fig. 11.

Table 7 ‘Types’ of Mind-Reading in the Academic Literature
Full size table
Fig. 11
figure 11

‘Types’ of Mind-Reading that are Claimed to be Possible Now

Full size image

4 Discussion

4.1 Overall remarks

In our study, we analysed 569 articles and identified an increasing academic interest in brain and mind-reading, as evidenced by Fig. 2. About 91% of these articles affirm and/or allude to the possibility of mind-reading –46% at present, while another 45% anticipate its future potential– as shown in Fig. 3. Our research underscores significant discrepancies within the academic discourse concerning the functionalities of neurotechnologies, their interpretation of the mind (mental lexicon), and the specific dimensions they aim to address. For instance, as illustrated in Table 4, although the verb “read” may partially align with terms like “decode” and “decipher,” it markedly diverges from verbs such as “extract,” “map,” “peer into,” and “access.” This divergence highlights a conceptual inconsistency in articulating the process of ‘reading X in the brain.’ In a similar vein, Fig. 11 depicts that the lexical choices employed to define ‘what it is that X reads in the brain’ lack scientific congruence. Terms such as ‘visual percepts,’ ‘hidden thoughts,’ ‘neural signals,’ and ‘semantic content’ exhibit minimal overlap in their conceptual framework, highlighting a disparate depiction and understanding.

Our detailed examination of technology mentioned in these articles shows fMRI as the predominant technology in a substantial majority of citations (Table 6), but also a significant rise in mentions of AI and BCI since 2019 (Fig. 7) which sharpens exponentially when brain-writing is correlated with AI and BCI (Fig. 5). Ethical issues are also featured prominently in these discussions; the analysis of articles connecting ‘mind-reading’ and ‘brain-reading’ showed an almost four-fold increase in the mentions of ethical considerations between 2012 and 2022. The most cited ethical consideration was mental privacy (30% of articles), followed by mental freedom (21%), personhood (11%) and mental integrity (10%) (Table 5).

In terms of mind-reading hype, fMRI was identified by 75% of articles acknowledging hyperbolic assertions, and this is consistent with its mention being on a downward trajectory. Accordingly, this trend in our research suggests that fMRI has moved beyond the summit of its inflated expectations about mind-reading (Fig. 10). However, even if AI, BCI and ANN are less frequently cited, it does not mean their assertions are not exaggerated claims. Our study shows that their increasing citation within a sample of academic literature indicates a trajectory toward the rising phase of inflated expectations (Fig. 8, 9 and 10).

4.2 Reading the mind as a metaphor

Our findings suggest a growing usage of the verb ‘read’ in conjunction with recent advancements in neurotechnologies, particularly where AI and BCI are increasingly utilised. However, claims and statements suggesting “reading x in the brain” are not a new phenomenon. The notion of reading the mind from the brain is a longstanding concept in Western scientific tradition [45] This idea evolved from the anatomists like Vesalius, who pioneered the readability of the body (the ability to read X from the body) to the metaphorical interpretation introduced by Galileo, who asserted that nature is written in the language of mathematics and, to fully grasp it, we must learn to decode its 'letters'. Taking this metaphor a step further, in 1755 the physician Guillaume-Lambert Godart was the first to explicitly suggest that one could 'read' the thoughts of a person using appropriate tools to interpret the 'letters' found in the brain [45]. In his treatise "La Physique de l’Âme Humaine" (The Physics of the Human Soul), Godart's central inquiry was to ascertain the corporeal localisation of mental faculties. He posited that each sensation and idea left a unique imprint upon the fibers of the corpus callosum, thereby suggesting a physical basis to be read with a “perfect microscope” [45].

These historical perspectives reflect an enduring belief or rhetorical aspiration for the readability of the body and brain, akin to interpreting a text. Indeed, if we were able to read one’s brain, does that not make us the equivalent of a book? Does that not reduce us to being no more than a thing to be read? The readability hypothesis suggests that a clear observation of the brain, either through Godart’s well-trained eyes or cutting-edge technology (e.g. AI, BCI, neurotech, etc.), would enable the deciphering of the brain's essence and its content.

As Aristotle noted, metaphors act as condensed analogies. He illustrates this explicitly in the Poetics (XXI, 1457b): “As old age is to life, so is evening to day. One will accordingly describe evening as the ‘old age of the day’ and old age as the ‘evening of life’” [94]. The structural relations in Aristotle’s example demonstrate a logic whereby “ ‘A is to B as C is to D’ yields the expression ‘C of B’ to designate A” [94]. In our context, we can see that: “(A) AI and BCIs are to (B) the brain, as (C) reading is to (D) symbolic expression (e.g., written language).” Accordingly, this yields the expression “reading of the brain” (‘C of B’) to designate (A) AI and BCIs.

As observed in our findings, many scholars now speak as if the newly demonstrated capacity of AI and BCI to capture brain activities will surely lead us to understanding the mind—as if the mind can be read directly and reducibly from the brain. The concept of mind-reading is an evolving facet of scientific aspiration and pretension. Its continuous and nuanced refinement reflects the rapid advancements in science. However, doing so increases the distorting effects of using metaphorical language to represent the progress in AI and BCI applications. Such linguistic choices often lead to conveying meanings that diverge from those intended in the original scientific publications (see section “Conclusions of studies extrapolated from mind-reading claims” below). This can introduce ambiguities and misinterpretations into the wider academic dialogue, particularly concerning the ethical considerations of mind-reading.

Recent usage of the readability of the brain metaphor in the context of AI and BCI as captured in our study does not mark a novel way of conceptualising and discussing the brain and its content. This metaphor frames the brain as an entity that can be read and understood. While treating the brain as a readable object by AI or BCI is not inherently misleading, it is crucial to explore not only the origins of this metaphor, as detailed above and elsewhere [45], but also to understand how, like many metaphors in science, it has the potential to both illuminate, obscure, oversimplify, hype or distort the actual findings.

4.3 Mind-reading in the scientific discourse

As shown by Table 7, there is no consensus in the academic literature about what neuroscientific mind-reading is. There are many different types of mind-reading studies used in the literature. To gain understanding, consider the experimental paradigms that have been frequently cited as empirical evidence supporting the assertions of mind-reading capabilities. These paradigms offer a comprehensive view of the methodologies and approaches employed in the research that underpin claims of mind-reading. For instance, the most cited evidence of mind-reading (68% of articles) was the ‘reading’ of a subject’s basic visual percepts from fMRI (e.g., object categories, orientation, and movement (see e.g. [24, 46]). Some articles (25%) pointed to studies in which researchers could ‘read’ more advanced images (e.g. natural images and ‘mind-movies’, (see e.g. [47, 48]) or semantic content (23%) (i.e. words or speech) (see e.g. [49]) held in working memory. Almost a third of articles (30%) suggested that neurotech could be used to ‘read’ hidden thoughts (i.e. decisions, intentions, and preferences) (see e.g. [50]) while for other authors (25%), just the neural signals picked up by BCI were enough to substantiate a mind-reading claim (see e.g. [51]). Claims were also made about neuroimaging enabling researchers to ‘read’ deceit (20%), emotion (14%), dreams (5%) or pain (4%). In short, while some authors use the mind-reading metaphor to describe something as basic as the detection of electrical signals in the brain using electrodes, others use it to claim sophisticated ‘decoding’ of our private intentions and beliefs using neurotech.

This brings us to observe that the diverse and often inconsistent application of the term mind-reading across numerous studies—with interpretations ranging from reading percepts to discerning preferences/intentions—suggests the absence of a robust, standardised definition in the scientific community. Our study highlights the imprecise and inconsistent usage of the term 'mind-reading' in scientific discourse, which points toward a lack of clarity and uniformity in understanding.

We understand that the varied interpretations of the term mind-reading demonstrate a reliance on traditional, intuitive, non-scientific interpretations of what mind-reading entails. When researchers select narratives about AI’s abilities that best explain their data, they are influenced by these preconceptions, and as such they are engaging in a form of informal descriptions of the findings. We contend that, in the absence of more precise explanatory models, researchers often resort to superimposing established terminologies on their data. This practice is evident across our study in the interpretation of mind-reading, whether it is conceptualised as accessing ‘hidden thoughts’ or discerning ‘emotional states’, thereby relying on the traditional readability narratives instead of a standardised, scientifically endorsed conceptualisation of mind-reading. To a certain extent, this appropriation of semantically laden terminology mirrors informal and subjective interpretations, potentially aligning with a paradigm of folk science and folk psychology. This is influenced by the concept of readability as a measurement that is implausibly objective. This approach in applying the term mind-reading without a universally acknowledged scientific foundation suggests that the subject of such study—the mind—is being interpreted through the lens of folk psychology. This metaphorical usage reflects a common, intuitive understanding of ‘mind’ rather than one that is firmly rooted in empirical evidence.

What is central about the notion of mind in the context of readability, in particular for ethics, it is that it involves a sense that our mind is private [52]. Our study shows that mental privacy was the highest rated ethical consideration amongst articles we analysed in our review, and it is a core concept in the discussion about neurorights [53,54,55,56]. Tong and Pratte’s [57] much cited work which characterises mind-reading as the decoding of information from the brain that is ‘fundamentally private and subjective’, conveys this meaning. This sense of mental privacy lies at the heart of AI and BCI concerns and prediction around mind-reading. Privacy can have descriptive or normative dimensions; it can relate to information or the body; it can concern accessibility or control; and it is culturally relative [58]. It’s therefore no surprise that there is no systematic science for describing the experience of human privacy [59]. Also, it is to be expected that speculation of its potential implications often leads to extrapolation beyond the current empirical evidence (see section "From AI hype to AI ethics hype: the case of neurorights" below).

4.4 Conclusions of studies extrapolated from mind-reading claims

Numerous narratives and reports concerning purported mind-reading technologies often reference primary studies that eschew the use of 'readability of the mind’ terminology in their own findings’ descriptions. The most recent example, at the time of writing this manuscript, in the influential Nature, was entitled: “Mind-reading machines are here: is it time to worry?” written by Reardon [12] which discussed, mostly, Tang et al.’s 2023 results [49]. In this study, Tang et al. present a non-invasive decoder capable of reconstructing continuous language from semantic representations in the brain, recorded via functional magnetic resonance imaging (fMRI). This decoder can produce sequences of words from brain recordings that may be intelligible or unintelligible, encompassing the exact meaning, the gist, or errors in the interpretation of both perceived and imagined speech, as well as silent videos.

It is noteworthy that Reardon's portrayal of Tang et al.'s findings categorises them as a form of mind-reading, despite the absence of the verb reading or the term mind in Tang et al.’s original article to describe their results. Adding to this observation, Reardon's article, as of December 2023, has been already cited six times in second and thirdhand academic literature, being referenced as scholarly evidence to substantiate the feasibility of mind-reading or read-out technology. However, it is critical to stress that Tang et al. do not characterise their findings as mind-reading and instead emphasise the necessity of subject cooperation for effective decoder training. This implies that in the absence of such cooperation, the decoding process is rendered infeasible, thereby opening avenues for potential circumvention of the decoding system:

“This demonstrates that semantic decoding can be consciously resisted in an adversarial scenario and that this resistance cannot be overcome by focusing the decoder only on specific brain regions” (Tang et al. 2023) [49].

The latter part of this statement is essential for understanding AI hype related to putative mind-reading ability: “resistance cannot be overcome by concentrating the decoder exclusively on specific brain regions.” This assertion is pivotal, as it explicitly delineates that the decoding of neural content cannot be achieved solely through the examination of brain regions or data. It implies that the efficacy of the decoding apparatus is contingent upon the subject's willingness to participate. As such, it stresses that a decoder, when trained on an individual's thought patterns, demonstrated inefficacy in discerning semantic details from another participant's data. As Tang et al. do not conclude or use the term mind-reading, it is strongly questionable to employ mind-reading to describe their research.

Describing firsthand studies in terms of mind-reading when these studies do not affirm such conclusions, points toward oversimplification and leads to an overly optimistic view of these technologies’ future capabilities, inadvertently fuelling hype. Our preliminary review of the scientific literature reveals an absence of evidence indicating proximity to achieving access to mental content without participant cooperation, notwithstanding the escalating prevalence of mind-reading terminology within scholarly discourse. Such practice, while not intended to mislead readers, nevertheless contributes to hyping AI capacities. Indeed, AI mind-reading capabilities remain markedly distant from science fiction narratives of accessing secretly hidden thoughts, that envision the deployment of brain scanners at borders, designed to detect deceit in individuals regarding their intentions, such as committing terrorist acts [60, 96]. It also demonstrates how inaccurate academic descriptions and poor referencing can contribute to the creation of urban legends [61].

4.5 Technological barriers to reading the mind

As reported above, the academic discourse includes assertions regarding neuroimaging's potential to 'read' deceit, accounting for 20% (Table 7). A substantial portion of scholarly debates has revolved around the feasibility of employing fMRI as a mind-reading tool for lie detection. Our analysis has identified a delayed speculative bubble surrounding fMRI's capabilities (Figs. 7 and 10), subsequently deflating due in part to the initial portrayal of fMRI as a means to ‘read hidden intentions’ and its purported efficacy in mind-reading lie detection. However, as empirical evidence accumulated over the years, highlighting the ineffectiveness of such technology, these claims began to diminish, particularly in light of studies demonstrating the inefficacy of neuroimaging-based lie detectors [62]. A notable feature of these now vanished claims is their focus on the individual’s control or consent in disclosing the content of their private and secret thoughts, a critical aspect in the ethical assessment of such technologies.

Our review has shown that scenarios in which individuals do have control over the sharing of their semantic data, which pertains to private cognitive states such as thoughts, opinions, and decisions, represent the type of violation of privacy by mind-reading which ethically matters the most [63, 64]. In the ongoing discussions about neurorights, the primary focus for protection centers on the non-consented [65,66,67] aspect of mind-reading, or veto control [64]. The ethical and legal implications of non-consented mind-reading scenarios receive significant scrutiny, although not all instances necessarily raise moral concerns. For example, intuitively inferring the thoughts of friends without explicit consent may not be ethically problematic [68]. Nonetheless, it is these types of hypothetical mind-reading scenarios—those conducted without consent by AI and BCI—that predominantly resonate within ethical and legal discussions.

Tang et al.’s assertion above delineates two fundamental constraints inherent in semantic decoding technology: (i) the absolute dependence on participant compliance, and (ii) the impasse in achieving generalisability across a diverse range of participants. These limitations crucially inform our understanding of the potential applications of mind-reading technologies, particularly in non-consented or non-control contexts where their use might be envisioned for concealing (e.g., unbeknownst to a subject) or enforcing (e.g., against someone's will) neural activity. While interesting and of great philosophical value [38], such hypotheses should be regarded as largely theoretical and reminiscent of speculative fiction. Tang et al.’s study demonstrates that without consent, control, or cooperation, it is technically impossible to achieve semantic decoding (e.g. via concealing and enforcing).

As delineated above, the objective of neurotechnological endeavours in decoding neural data is to correlate specific functions and processes with distinct brain regions; however, it cannot do so without intensive personalised (and non-generalisable) AI training. The idea that reading neural data content—the mind—only requires neurotechnological access to certain brain regions points toward fiction rather than science. This fictional functionalist and reductionist account encounters challenges in encompassing the nuances of conscious experience [69, 70]. Moreover, even assuming functionalism aligns with colloquial understandings of the mind, the task of precisely mapping its constituents onto specific brain regions is beyond complex, and so far, appears un-replicable from individual to individual. This complexity highlights the inherent limitations of AI and BCI reductionist methodologies in effectively interpreting the mind's readability.

Additionally, there are conceptual limitations of mind-reading endeavours relating to their reliance on correlation and inference. We cannot overcome the fact that neuroimaging data are inherently correlational [71]. That is, mind-reading studies can only demonstrate statistical correlations between patterns of brain activity and mental information, and correlation is not causation [72]; just because mental information can be decoded with a certain measure of AI predictability from representations of brain activity, like in Tang et al. [49] study, it does not mean that there is a causal relationship. Further, being able to decode information from patterns of neural activity does not give us reasonable grounds to infer that this information is neurally represented or functionally exploited by the brain [71, 73, 74]. This is more than an objection to making inferences from information to representation. Current neurotechnologies like fMRI, AI and BCI can at best show that there is latent information about experimental conditions in neural activity. This does not mean that the brain uses this information, or even that the information is usable [74].

More generally, the use of inference complicates the conclusions that can be drawn from mind-reading studies. Issues arise when deductive claims are made from studies using the common technique of 'reverse' inference. In logic terms, this is the fallacy of affirming the consequent [75]. This is not to suggest that reverse inference is not a useful means of scientific reasoning as an example of ‘abductive inference’ or ‘reasoning to the best explanation’; however, it becomes problematic when the mind-reading studies’ hypotheses are reified as facts [71].

The misrepresentation of study results is of particular concern because the inferences that are justified by our knowledge of brain function are very coarse [76]. The most-cited mind-reading studies [see., eg 47], establish relationships between stimuli and brain activity patterns. Although some mind-reading claims in the literature suggest otherwise, no experience-specific brain activity patterns have been established, and given the brain’s non reductive nature, this is very unlikely to happen [77]. Decoding results should therefore be interpreted and communicated with care and consideration should be given to augmenting these results with behavioural models [73] or other self-reported data. This may help to counter the misguided perception that neurotechnologies like fMRI, AI and BCI can ‘read’ the mind.

4.6 From AI hype to AI ethics hype: the case of neurorights

In science, hype can lead to a peak of inflated expectations, where the initial excitement is disproportionate to the actual capabilities of a technology [78]. For our study, we understand that hype refers to the enthusiastic discourse surrounding AI and BCI capacities to read the brain, often leading to exaggerated expectations of reading the mind. Our study shows a clear example of brain-/mind-reading hype with fMRI (Fig. 7, 8, 9, 10). Indeed, 21% of the relevant articles suggest a certain level of hype around mind-reading claims. FMRI, implicated in 75% of these claims, shows a decreasing trend in mentions, indicating the likely move past its hype phase. In contrast, mentions of BCI, AI, and ANN are on the rise, suggesting they are in the ascending speculative phase of heightened claims and expectations (Fig. 10). Are we just seeing the beginning of the hype for BCI and AI?

As demonstrated in our study, putative AI-enabled brain-reading neurotechnologies for mind-reading purposes were associated with ethical concerns such as violation of mental privacy, mental freedom and personhood. In the realm of AI technology, there exists a proportional relationship between AI and ethical hype. As technological claims become more inflated, so too do the corresponding ethical concerns. Ethical AI hype specifically refers to the fervent and sometimes overstated promotion of ethical risks associated with AI technology. It is predominantly characterised by speculative ethics [79, 80]. In terms of neurotechnology, this form of speculation tends to emphasise potential problems that may never materialise, leading to exaggerated claims of potential threats to notion such as free will [81, 82].

Speculative ethics involves hypothesising about various outcomes or possibilities without conclusive evidence. It typically leans towards forecasting future scenarios based on current trends, data, and observations, albeit lacking definitive evidence, proof or certainty. As indicated in our study, a significant portion of the literature—45% of articles—projected that mind-reading would be feasible in the future, while 46% claimed it was already achievable, without much evidential substance. This distribution underscores the speculative nature of ethical discourse in this field, illustrating how projections about future capabilities can shape ethical discussions, often without a solid foundation in the current AI and BCI realities.

While our study shows evidence of AI and BCI hype in relation to putative claims of AI-enabled brain-reading neurotechnologies for mind-reading purposes, it finds the presence of an AI ethics hype too. This is not new: most emerging technologies are associated with a bubble of inflated ethical scrutiny. Occurrences of “ethics hype” and “speculative ethics” are also present in other domains of ELSI literature [79, 80, 83,84,85,86,87,88,89,90].

Anticipating the potential future challenges of AI's ability to interpret brain activity is a crucial aspect of AI ethics. However, this forward-looking approach can sometimes lead to fruitless speculation on improbable scenarios, giving rise to redundant ethical prescriptions. Speculative ethics is linguistically marked by hypothetical reasoning evident in phrases like “suppose that…”, “what if…?”, and “if… then.” Alfred Nordmann [79] characterised this as the “if and then” syndrome, where speculative statements begin by positing potential future technological developments (the “if”), followed by presumed inevitable outcomes (the “then”) [86]. For example: “if we can read and transcript neural activity, [then] we might be able to read and transcript minds”. [70, 91] Such an example clearly indicates that a hypothetical and improbable technological scenario “read and transcript minds” ineluctably derives from the false equivalence of “read and transcript neural activity” [91]. As such, it allows an imagined future to overshadow the present [79].

Our review of the literature has found that in the realm of AI and its prospective issues, the term “therefore” was added to this “if and then” construction [80]. This addition led to that type of logical structure (a) if, (b) then; (c) therefore [80]. For instance, a logical statement such as: (a) if AI can interpret visual perceptions from the brain, (b) then it might hypothetically soon lead to reading hidden thoughts of the mind; (c) therefore, we urgently need neurorights to safeguard mental privacy. For instance, let us take the most recent illustration suggested in the literature (Nov 2023):

“the possibility of reading thoughts […] could theoretically be used to reveal people’s thoughts without their consent, and even for malicious purposes, such as blackmail and discrimination […] the formal recognition of a right to mental privacy, as proposed by several authors […] could contribute to mitigating the misuse of mind-reading technologies” [56].

Although the hypothetical syllogisms do not explicitly employ “if–then-therefore” terminology, it is clear that the reasoning is structured conditionally (a), (b); (c), resulting in an inflated prescription. Again, while such examples are ethically and philosophically worthy of exploration, and are not intended to mislead readers; nonetheless, they contribute to inflating and distorting AI capabilities by incorporating the concept of mind-reading. Our review revealed that the ‘we-need-neurorights-to-protect-us-from-mind-reading’ hypothesis points toward imaginative reasoning based on improbable technological outcome. It posits that if mind-reading were to become a reality, dystopian scenarios would likely ensue, necessitating proactive implementation of protective measures. We have found that the “delay fallacy” [38, 92] likely underlies the inception of these categories of AI mind-reading hype about neuroright statements. This type of if–then-therefore approach, especially when applied to the promising field of AI and reliant on prophetic claims, increases the risks of inflating this aspect of AI ethics to merely a narrative exercise rather than preserving it as a rigorous scientific discipline.

5 Limitations of the study

A potential limitation of this study, which paradoxically reinforces our central argument, lies in our methodological choice to specifically search for terms like ‘brain-reading’ and ‘mind-reading’ within the academic literature. On one hand, had our search been broader, encompassing terms such as “brain decoding”, “mind decoding” we might have encountered a more extensive corpus of articles. Indeed, a cursory search using these expanded terms resulted in approximately 3500 entries, indicating that our chosen keywords may not have fully captured the entire spectrum of ‘brain-reading,’ a concept more expansively denoted by ‘brain decoding’.

However, while Google Scholar is a powerful tool, it’s important to note that relying on any single source could potentially limit the scope and depth of our analysis. Diversifying our sources could facilitate a more comprehensive and nuanced understanding of the discourse studied. This reinforces the need for a more expansive and diversified search strategy in future studies to capture the full breadth and depth of the discourse on ‘brain-reading’ and ‘mind-reading’.

6 Reflectivity on the study

While qualitative analysis can highlight the discursive construction of ethical issues through articles, it is inherently interpretive and may be susceptible to theoretical bias [21]. The definition of suitable evidence can vary depending on a researcher’s theoretical perspective. Therefore, a significant challenge in our study was determining what constitutes evidence and how to interpret that evidence to demonstrate objective findings.

Our initial task was to look at ‘mind-eading’ and ‘brain-reading’ based on preliminary work done by one of the authors [45, 95]. This current work has led to an incidental discovery of hype within the exiting literature, and an incidental finding of increased mentions of brain-writing. While our study was not initially crafted to specifically track claims pertaining to neurorights, it quickly emerged that there was a robust correlation between the variables of ‘AI’, ‘mind-reading’ and ‘neurorights’. Had we focused on these search terms, we likely would have had a better opportunity to investigate further claims surrounding neurorights related to mind-reading, potentially demonstrating further evidence of hype.

Given our results partly originate from intricate interpretative processes, it was crucial to employ a balanced mix of quantifiable data and interpretation [21]. This approach, which grounds interpretations in the academic literature, was necessary to mitigate potential bias. It ensures that our analysis was not only nuanced but also anchored in empirical evidence. In essence, the combination of quantifiable data and interpretation, along with a careful choice of search terms, helps mitigate the risk of bias [21]. This approach underscores the importance of methodological rigor in our study which includes complex concepts such as ‘mind-reading’ and ‘neurorights’.

Conversely, and as we have previously articulated, the term ‘mind-reading’ bears an inherently non-scientific connotation. While empirical studies employing the notion of ‘mind-reading’ do exist, they frequently fail to encapsulate the philosophical depth intrinsic to this concept. The usage of ‘mind-reading’ imbues the research with an element of spectacle, generating a profound impact that more neutral or scientifically precise terms may not achieve.

7 Conclusion

Our interpretive content analysis encompassed an analysis of 1017 articles published between 1972 and 2023, focusing on the phenomena of mind-reading and associated claims. During this period, we observed a decline in the hype surrounding fMRI, characterised in part by assertions of its utility in mind-reading for lie or terrorist detection [60, 96]. Our findings suggest that we are currently witnessing the nascent stages of heightened interest and speculation concerning BCI and AI, with a focus on non-consented access to private thought, thereby requiring neurorights for protection.

The term ‘mind-reading’ within academic discourse is employed with significant variability. Some researchers use it to refer to basic detection of electrical brain signals via electrodes, while others extend its meaning to the intricate process of ‘decoding’ private intentions and beliefs using BCI and AI techniques. This variation in terminology, while not necessarily deceptive within specific scientific circles, can lead to misconceptions about neurotechnological capacities in more interdisciplinary applications, particularly in the realms of ethics. This highlights the critical need for precise and clear usage of scientific terms across different disciplines; as a multidisciplinary definition and understanding.

Moreover, an inevitable tendency towards oversimplification exists in these studies. It is crucial to differentiate between ‘brain-reading’ and ‘mind-reading,’ particularly given the ethical and philosophical concerns raised. Few critiques or studies address the exaggerations often associated with these claims [81, 93]. The forms of mind-reading that carry substantial weight in ethical discussions involve scenarios where individual consent and control are pivotal factors [63,64,65,66,67].

Our results point towards the formation of a rhetorical bubble, encompassing both scientific and ethical narratives, fueled by the burgeoning hype around AI. Notably, our study protocol was not initially designed to track claims related to neurorights. However, a strong correlation between the use of 'mind-reading' and the discussion of neurorights emerged during our investigation. While imaginative scenarios derived from mind-reading neurotechnology can stimulate engaging ethical and philosophical debates, they can also divert research into unproductive and speculative areas, potentially impeding neuroethical progress [79, 80, 83]. Despite the hype surrounding some aspects of neurorights, particularly regarding mind-reading claims, the necessity for stringent regulations and policies governing AI and BCI advancements—especially their impact on agency, access, consent, privacy, and data extraction—remains crucial and should not be overlooked [54, 97,98,99,100,101,102,103,104,105,106]. Future research is warranted in this area to elucidate the hype surrounding neurorights associated with mind-reading claims.