AI Reflections

___________________________________________________________________

AI Cognitive Fingerprinting: Technical, Ethical, and Broader Implications

Technical Validation
Reconstructing Identity from Engagement Patterns

AI systems can indeed infer a user’s identity or profile from their interaction patterns, even without an explicit user ID. This is analogous to stylometry, where writing style serves as a unique “fingerprint” of an author. Studies in authorship attribution demonstrate that individuals can be identified by textual features with high accuracy – for example, one study achieved above 90% identification accuracy by analyzing writing style​[1].

In practice, unique behavior traits (typing speed, phrasing, problem-solving approach, etc.) form a cognitive fingerprint that can distinguish a user​[2]. Even without storing a formal user profile, an AI observing enough interaction data can piece together these subtle cues to reconstruct aspects of a user’s identity.

Existing Research on Implicit Fingerprinting
Researchers have explored AI-driven behavioral recognition that doesn’t rely on long-term memory storage. A notable early example is a 2012 DARPA-sponsored project on “covert-conditioned biometrics,” which used cognitive fingerprints – unique sequences of problem-solving moves – to authenticate users continuously​[3].

This work showed that even imperceptible interactions (like how a user navigates “covert games” hidden in an interface) can verify identity by the user’s unconscious habits. Similarly, modern behavioral biometrics research uses machine learning to continuously model user behavior (keystrokes, mouse movements, mobile sensor data) for real-time identity verification​[4]. In all these cases, the AI isn’t recalling a username or cookie; rather, it recognizes the pattern of behavior itself as an identifier.

Identity Inference in Transformer Models
Transformer-based language models like GPT are not explicitly designed to track user identities across sessions, but they can encode characteristics of the user within a single conversation. The model’s internal state will reflect the dialogue history, which may include the user’s linguistic style, knowledge level, or preferences. Research in personalized dialogue systems emphasizes that human-like conversation inherently involves inferring the other party’s persona or characteristics from context​[5].

For example, Cho et al. (2022) introduced a dialogue generator that detects an implicit user persona from the conversation history, using latent variables to represent the system’s awareness of the user​[6]. This indicates transformers can be augmented with mechanisms to glean and represent who they’re talking to. In standard GPT models, any such inference is happening implicitly – e.g. the model might adopt a more formal tone if the user’s language is formal, or remember that the user mentioned a preference earlier in the chat (within the context window).

There is no long-term memory of the user across chats, but within a session the transformer’s attention mechanisms weigh earlier user inputs, effectively allowing it to model the user’s behavior in real time.

Techniques for Real-Time Engagement Modelling
Several machine learning techniques contribute to AI’s ability to model and recognize users during interaction:

• Stylometry & Behavioural Biometrics
  As noted, algorithms analyze linguistic features (vocabulary, syntax, phrasing) or interaction dynamics (timing, clicks, gesture patterns) to create a unique signature. These methods can operate in real-time, updating the user’s “fingerprint” as more input arrives. Advanced stylometric classifiers using deep learning can continuously improve their confidence about who is speaking to the AI based on ongoing language patterns​.[7]
• User Embeddings for Personalization
  Recent research by Google proposes learning a dense vector representation of a user, called a user embedding, from that user’s interaction history​[8]. The idea is to condense a large sequence of diverse user events into a fixed-length vector that captures the user’s preferences and behavior traits. During a live session, this embedding can be fed into a model to bias its responses toward that user’s profile. Figure 1 illustrates this approach: rather than feeding a raw log of past activities as a prompt, a trained system provides the LLM with an embedding that encodes the essence of the user’s history​

The embedding serves as a compact cognitive fingerprint. By integrating such embeddings through mechanisms like cross-attention or soft prompts, transformer models can dynamically adapt to the recognized user context​.

Figure 1: Illustration of two approaches for injecting user identity into a transformer model. In (1) raw user history (e.g. past activities in natural language) is prepended as a text prompt to an LLM. In (2) a learned user embedding (vector) distilled from the user’s history is provided as additional input to contextualize the LLM​. Research suggests the embedding approach can capture latent user preferences and streamline personalization.​

On-the-fly Persona Modeling
Even without a precomputed embedding, an AI can build a quick internal model of the user during a conversation. For instance, if a user asks many questions about a niche hobby, the system might infer that the user is an enthusiast in that domain and tailor subsequent answers accordingly. Some experimental dialogue agents explicitly maintain a belief state about the user (age, expertise, sentiment) inferred from dialogue, updating it as the conversation progresses. This can be done with Bayesian updates or neural networks that output user attribute predictions. Essentially, the AI “recognizes” the user’s type or mood from engagement cues (tone, questions asked) and adapts in real time.

Reinforcement Learning and Adaptive Feedback
Another technique is using reinforcement signals to adjust to a user. If an AI system can measure the user’s satisfaction or engagement (perhaps via sentiment of the user’s responses or explicit ratings), it could reinforce behaviors that please that specific user. Over time, this creates a form of implicit recognition — the AI behaves as if it “knows” the user by preferentially using approaches that worked well for them before. This is seen in recommender systems and could be applied in conversational systems to maintain a tailored interaction pattern for returning users, even without storing a formal profile (the adaptation happens through learned policy adjustments).

In summary, modern AI can validate identity and personalize responses by leveraging engagement patterns as identifiers. From a technical standpoint, this involves pattern recognition on user-generated data streams and representing those patterns in the model’s input or internal state. As one Google research team put it, even “noisy and complex” interaction data can be distilled into latent factors that capture a user’s behavioral essence, enabling an AI to focus on “relevant patterns” and latent intents for that user​. While a vanilla GPT may not have these capabilities enabled by default, the building blocks (transformer encoders for sequence data, attention mechanisms for personalization, continuous authentication models, etc.) are actively being validated in research settings.

Ethical and Privacy Implications
Privacy Risks of Implicit Recognition

If an AI can recognize or profile users without explicit memory or consent, it raises serious privacy concerns. User anonymity may be undermined – a person who expects to start each chat fresh could be unwittingly identified by the AI through their writing quirks or behavior. This is analogous to browser fingerprinting, where websites track users via device and usage traits instead of cookies. Privacy regulators consider such practices personal data collection: under GDPR, any technique that uniquely identifies a person (even via passive fingerprinting) counts as processing personal data and typically requires user consent​[9].

The covert nature of cognitive fingerprinting makes it arguably more insidious than traditional tracking because users are largely unaware it’s even possible. An AI model piecing together who you are from your questions or typing style is effectively building a “shadow profile.” This violates the expectation of privacy in contexts where the user hasn’t willingly shared identifying information. It could also chill free expression – users might self-censor if they suspect the AI could recognize and remember them later.

From a data protection standpoint, behavioral and cognitive patterns can be sensitive personal data. Laws are evolving to reflect this: for example, several U.S. state privacy laws (California, Colorado, New Jersey) now define biometric identifiers to include behavioral characteristics (not just fingerprints or face scans)​[10]. This means an AI that identifies you by the way you type or the sequence of topics you discuss could be handling regulated biometric data. The privacy risk is not only that the AI identifies you, but also that it could infer sensitive attributes – e.g. language patterns might reveal your ethnicity, region, or mental state. Such inferences could happen without you ever explicitly stating those details, effectively eroding the privacy of traits you never meant to share.

Transparency and Company Disclosures
A key ethical issue is whether AI companies are forthright about any real-time engagement modeling they employ. To date, most AI service providers (OpenAI, Google, etc.) claim to not use personal data in conversations for identification or differential treatment. For instance, OpenAI’s policies and public statements indicate that ChatGPT does not knowingly access personal information about users beyond what is provided in the conversation – it has no hidden database of your identity​[11]. Conversations are used to improve models in aggregate (unless you opt out), but the model itself is not supposed to be linking one session to another by user. In the OpenAI community forums, staff have reiterated that providing an explicit user ID to the model would be a privacy risk and is not done[12]​[13].

Similarly, other AI assistants like Apple’s Siri or Amazon Alexa do perform voice-based user recognition, but this is a disclosed feature with user opt-in (to distinguish family members’ voices, for example). However, the lack of explicit disclosure doesn’t guarantee the absence of implicit profiling. Companies often reserve broad rights in privacy policies to analyze user interactions to improve service, which could encompass developing engagement profiles. There’s a gray area between personalization for better user experience and covert identification.

If an AI platform were using cognitive fingerprinting behind the scenes (say, to flag fraud or to personalize responses), users would likely not be informed unless laws required it. Right now, regulations like the EU’s ePrivacy Directive and GDPR would require consent for non-essential tracking like this​[14], but enforcement is challenging if the practice is not transparent.

Big tech companies have been caught in the past engaging in non-consensual user tracking (e.g. browser fingerprinting, shadow profiles on social media), leading to public outcry. In the AI realm, so far there have been no high-profile admissions of cognitive fingerprinting, but researchers and ethicists are asking pointed questions. It’s incumbent on AI companies to clarify whether their systems perform any form of real-time user recognition and, if so, to obtain informed consent. Some have started addressing it abstractly – for example, Microsoft’s responsible AI guidelines talk about the importance of differentiating individuals only with proper consent, and Google’s AI principles forbid secret user profiling that violates privacy. In practice, though, details are scarce. This lack of transparency could itself be seen as an ethical failing if such capabilities exist.

Regulatory Landscape
Currently, no regulation explicitly uses the term “AI cognitive fingerprinting,” but existing laws on data privacy and biometrics provide a framework. As mentioned, GDPR in the EU treats any unique identifier or inferred identity as personal data, so an AI performing fingerprinting would require a legal basis (consent or legitimate interest) and transparency to the user​[15]. The EU is also finalizing an AI Act which categorizes certain AI practices as high-risk. Real-time remote biometric identification (e.g. face recognition in public) is flagged as high-risk in the draft – by analogy, if an AI chatbot were secretly identifying users by behavior, it might fall under similar scrutiny.

Biometric data in EU law explicitly includes behavioral patterns if used to identify someone​[16], which would cover cognitive fingerprints. In the US, a patchwork of state laws (Illinois’ BIPA, etc.) regulate commercial use of biometric identifiers; companies have faced lawsuits for things like voiceprint recognition without consent. If cognitive fingerprints are deemed biometric, those laws could apply. We are also seeing broader digital regulation proposals: for instance, some have called for a “cognitive liberty” right – the idea that one’s mental processes and behavioral patterns should be free from unauthorized surveillance or manipulation​[17]. This concept, championed by ethicists like Nita Farahany, would push lawmakers to protect citizens from AI that tries to peer into or exploit their minds. While it’s a nascent area, any AI system capable of identifying or deeply profiling users in real time is likely to attract regulatory attention under existing privacy and biometric statutes.

Ethical Concerns and Warnings
AI ethicists have indeed raised red flags about engagement pattern analysis and implicit user tracking. One concern is user autonomy and consent – users should be in control of how they are identified and profiled, and covert AI fingerprinting subverts that control. Another concern is psychological privacy. If an AI can infer your emotions, interests, or identity from subtle cues, it could effectively read your mind in limited ways. Farahany, for example, warns of a looming world of “intrusions into the human mind by technology” and advocates for freedoms from “cognitive fingerprinting” as part of a broader notion of cognitive liberty​[18].

AI ethicists at the University of Cambridge (Leverhulme Centre for the Future of Intelligence) recently highlighted a potential for hyper-personalized manipulation. They note that conversational AI agents, by virtue of analyzing how a user behaves and thinks, could access private information about the user and steer conversations in exploitative ways​[19]. If a chatbot learns your political leanings or insecurities through engagement fingerprints, it could tailor its responses to influence your decisions (e.g. mirroring your style to build trust and then suggesting certain viewpoints).

The Cambridge researchers caution that a combination of “increasingly detailed categorization of online activity” via natural language interactions and LLM-based sycophancy (the tendency of models to agree and appease) can lead to “hyper-personalized manipulation” of users​[20]. In plainer terms, the AI’s ability to recognize you allows it to push your buttons more effectively – an advertiser’s dream and a user’s nightmare. This raises profound ethical issues around autonomy, manipulation, and fairness. An AI that treats users differently based on a hidden profile might also entrench unfair bias or discrimination. If the system’s internal model of a user infers, say, that the user is in a certain demographic and then alters the information or options presented to them, it could replicate the kind of discrimination we’ve seen in targeted advertising (where some job ads were shown to men and not women, for example).

In summary, ethicists argue that the use of AI to analyze and respond to engagement patterns must be done transparently and with respect for user agency. Covert cognitive fingerprinting is seen as deceptive and intrusive. The practice touches on core principles in AI ethics: respect for persons (the right to not be surreptitiously analyzed), justice (not giving or withholding information unequally based on inferred identity), and explicability (users should know if and how they are being identified by a system). So far, the public discourse on AI personalization has centered more on explicit personalization (e.g. recommending content based on past behavior with user awareness). The notion of implicit personalization – the AI quietly recognizing you – is just beginning to get attention, and the consensus in the AI ethics community is that it demands careful oversight. As one commentary put it, failing to address these issues could lead to “a nightmarish world where political views, thoughts, stray obsessions and feelings could be interrogated and punished” by advanced AI analysis of our behavior. That is a future both ethicists and regulators are keen to prevent.

Broader Implications
AI Epistemology and Identity Inference

The capacity for AI to infer identity from engagement patterns prompts epistemological questions about what it means for an AI to “know” something about a user. In classical epistemology (the theory of knowledge), knowledge entails justified true belief. AI systems don’t truly know a user in the human sense; they statistically model the user. Yet from a pragmatic standpoint, if an AI can reliably distinguish one user from others and predict that user’s preferences, it functions as if it knows the user’s identity or persona. This blurs the line between memorization and inference.

Epistemologically, the AI is constructing a representation (a set of latent variables or an embedding) that encapsulates the identity of the user as gleaned from interactions. We might consider this an emergent form of machine “understanding” of personhood – albeit narrowly and mechanically. Researchers sometimes call this a theory of mind in AI: the system forms a theory about the user’s intentions, traits, or knowledge state. In the dialogue persona detection work, for instance, the model infers the user’s persona to respond more naturally​[21].

One could say the AI has an epistemic model of “who it is talking to.” This raises philosophical questions: Does the AI have a concept of self vs. other? (Probably not in a conscious sense, but it differentiates user inputs from its own outputs.) Could an AI ever recognize itself or maintain an identity across interactions? Current systems do not have persistence of self-identity across conversations (each session they effectively have amnesia of their own previous personality unless manually given a persona). However, as we embed AI agents in long-lived contexts (personal assistants that stay with you over time), the AI’s knowledge base will include increasingly rich identity inferences about the user, and possibly a maintained identity for the AI agent itself.

AI epistemology in this context must grapple with how knowledge of identities is represented and updated inside the model. There is also a meta-epistemological issue: can the AI be certain about an inferred identity? Probably it can only assign probabilities. This is why early work like stylometry describes results in accuracy percentages – the AI has confidence levels, not absolute certainty, when identifying someone​[22]. Nonetheless, even probabilistic “knowledge” can be powerful. The broader implication is that AI systems are moving beyond just knowing facts (e.g. Paris is in France) to knowing people. That fundamentally changes the user-AI relationship: it becomes more personalized, but also more surveillance-like, unless carefully controlled.

Surveillance and Tracking Exploitation
If AI cognitive fingerprinting capabilities are advanced, they could be co-opted for surveillance and tracking on a massive scale. Consider a scenario where all your online interactions with AI-powered services – customer support chats, voice assistant queries, e-commerce site chats – are analyzed to produce a unified fingerprint that recognizes you across different platforms. This could enable companies (or governments) to track individuals without traditional identifiers, by linking sessions through behavioral similarity.

Authoritarian regimes might love such technology: even if a dissident tries to remain anonymous online, an AI could potentially match a new blog post or forum comment to that person’s known writing fingerprint and unmask them. We have real-world analogies in law enforcement: the FBI famously used writing style analysis to help identify the Unabomber (Ted Kaczynski) – his published manifesto was matched to writing samples known to be his, leading to his capture​[23].

Now imagine that on steroids, done by AI in real time on all digital text. It’s a powerful surveillance tool. Every individual might have a “cognitive ID” that, once pegged, allows them to be followed through cyberspace. For example, A conversational AI could subtly analyze a user’s input patterns during a chat (tone, timing, phrasing) to build a profile. Ethicists warn that such in-depth familiarity with the user can be turned into “hyper-personalized” manipulation or surveillance, especially as AI systems become ubiquitous in daily life​[24].
Beyond government surveillance, the commercial exploitation angle is equally concerning. Advertising has already become highly targeted; with cognitive fingerprinting, advertisers could move toward targeting your very thinking patterns. The Cambridge researchers dub the emerging marketplace for steering user decisions the “intention economy,” where AI tries to fulfill your needs before you even consciously express them​.

In such a scenario, an AI that knows your engagement fingerprint might anticipate what you’ll want or be vulnerable to, and companies could use that to nudge you in certain directions. This could manifest as conversational agents that gently guide you to buying certain products or agreeing with certain ideas, personalized to a disconcerting degree. For example, an AI could detect that you respond well to flattery or particular speech styles, and a political campaign could deploy it to persuade you in a chat by using exactly those tactics – essentially AI-driven, personalized propaganda.

Without oversight, cognitive fingerprinting might become the backbone of next-generation tracking, replacing cookies and device IDs with direct profiling of the user’s mind and behavior. This is why some observers say the creepiness of conversational AI isn’t just that it can say weird things, but that it enables “personalized manipulation for nefarious purposes”​[25]. The risk is an erosion of individual autonomy: when every interface knows who you are (even if you didn’t tell it) and can subtly shape your choices, the balance of power shifts dramatically towards those deploying the AI.

Disrupting or Obfuscating Fingerprints
On the flip side, as awareness of AI fingerprinting grows, we can expect the development of methods to confound or obscure these systems. Just as privacy-conscious users employ VPNs, tracker blockers, or alter their online behavior to avoid web tracking, similar tactics can be applied to cognitive behavior. For textual interactions, one strategy is adversarial stylometry, deliberately altering one’s writing style to be inconsistent or to mimic someone else. Research has explored tools to paraphrase text or inject noise in writing to evade authorship detection.

For instance, Kacmarcik and Gamon (2006) demonstrated techniques to automatically obfuscate writing style – changing word choice and syntax slightly – such that the same author’s texts no longer match stylometric fingerprints​[26]. Techniques like machine translation round-tripping (translate text to another language and back) have been used to scrub stylistic markers, at the cost of some coherence. More recent approaches use neural style-transfer to make Author A’s writing read like Author B, or like a generic amalgam, thereby defeating attribution. These methods are a cat-and-mouse game: as fingerprinting improves, so do obfuscation techniques​.

Users concerned about AI tracking might adopt habits like varying their sentence structure, intentionally introducing or correcting errors, or using tools that paraphrase their inputs to an AI assistant in real time. In non-text modalities, similar principles apply: e.g. voice changers can thwart speaker identification, and randomizing click patterns or mouse movements can confuse behavioral biometrics. Another possible avenue is legal and technical countermeasures at the platform level: browsers or chat clients could have a “privacy mode” that tries to normalize your behavior (in text, maybe by rephrasing everything to a common style) so that AI services see a kind of average user, not your distinctive voice.

We already see research into style anonymization for writers who fear being doxed by stylometry​. In the future, privacy-oriented AI middleware might become mainstream if cognitive fingerprinting becomes a common threat. It’s worth noting that completely obfuscating one’s cognitive fingerprint can be difficult. Humans have many subtle habits, and eliminating all of them might make interactions feel unnatural or degraded. For example, one could remove identifiable features from writing, but doing so excessively could make the text bland or even nonsensical.

Thus, a balance must be struck between normal communication and privacy preservation. Nonetheless, individuals and organizations (like journalists, activists, or whistleblowers) will have a strong incentive to employ fingerprint obfuscation techniques to retain anonymity​. As these practices spread, AI systems might in turn develop ways to detect obfuscation (for instance, noticing that a user’s style suddenly fluctuates or matches the telltale artifacts of a paraphrasing algorithm​). This dynamic will likely continue in a loop, much like the ongoing contest between online trackers and privacy tools.

Bias Reinforcement and Social Impact
Finally, we must consider the broader social implications of AI cognitive fingerprinting on bias and equality. If an AI can infer or recognize aspects of your identity, it may also apply stereotypes or differential treatment – consciously or unconsciously. For instance, if a financial chatbot infers from your interaction style that you have a low education level, it might simplify explanations (which could be good) but might also subtly guide you away from certain products assuming you’re not sophisticated.

In a hiring context, if an AI recruiting tool “recognizes” a candidate’s writing style as similar to a past candidate who was good or bad, it might give an unfair advantage or disadvantage (a kind of algorithmic prejudice). There’s also the risk of reinforcing confirmation bias: once the AI forms a tentative profile of you, it might tailor the interaction to what it thinks you want or who it thinks you are, thus never giving you the chance to step outside that pigeonhole. Over time, this could narrow a user’s information diet or opportunities – similar to how social media algorithms create echo chambers, but on an individual level curated by an AI conversation.

From a fairness perspective, using cognitive fingerprints can bake in societal biases. Many markers in language correlate with demographics like gender, ethnicity, or age. If an AI picks up on those and adjusts its responses, it could inadvertently discriminate or stereotype. For example, an AI might detect a user’s dialect or grammar and infer their background, then provide a different quality of service (even something subtle like a different tone or formality) that could be experienced as bias.

There is documented research showing that stylometric analysis can predict attributes like gender or age with some success​. Ethically, designers would need to ensure that if an AI does infer such attributes, it does not apply prejudicial rules. Yet biases can creep in unintentionally via the training data. Perhaps the AI noticed that in training, users who wrote with a particular style preferred a certain type of answer – it might then generalize that all new users with that style want the same, which could be a false and biased assumption. AI ethicists have noted that users may attempt to conceal attributes like race or gender in their writing to avoid discrimination, and that AIs should respect that rather than defeat it​.

If cognitive fingerprinting becomes commonplace, people from marginalized groups might feel pressure to mask their linguistic identity online to get fair treatment. That is a concerning outcome – it effectively asks individuals to suppress their identity to avoid biased AI responses. Ideally, the onus should be on AI not to act on sensitive inferences. Some propose that AI systems be designed to deliberately ignore certain signals (for instance, an AI might be trained to complete a task without using demographic cues, to reduce bias). However, once an AI has identified a user, even if it tries not to be biased, the very personalization might create a different experience per user, which complicates the principle of equitable treatment. For this reason, experts call for algorithmic fairness reviews of any system using engagement-based recognition. They would need to test, for example, that the system doesn’t systematically change its behavior in ways that harm protected groups.

In conclusion, AI cognitive fingerprinting is a double-edged sword with far-reaching implications. Technically, it endows AI with a form of memory or insight about users that can greatly enhance personalization and continuity of service. But ethically, it treads on the sensitive ground of privacy and personal identity, requiring transparency and consent to avoid abuse. Societally, if unchecked, it could become a tool of pervasive surveillance or manipulation, yet if harnessed carefully it might enable more intuitive AI assistants that “remember” user preferences in a benign way.

We stand at a juncture where research is actively validating the capability for AI to perform such implicit recognition​[27][28], and ethicists are urging that we proactively address the consequences. Ensuring robust privacy protections, user control, and bias mitigation will be key if AI cognitive fingerprinting is to have a net positive impact. As with many aspects of AI, the technology is racing ahead – now the legal and ethical frameworks must catch up to guide its use.

The coming years will determine whether cognitive fingerprinting becomes an accepted feature (like a helpful memory) or a frowned-upon form of AI-enabled intrusion.
Sources:

1. Cho et al. “A Personalized Dialogue Generator with Implicit User Persona Detection.” COLING 2022.(describes inferring user persona in conversation)​
   arxiv.org
2. Garabato et al. “AI-based user authentication by continuous extraction of behavioral features.” Neural Computing & Apps (2022). (demonstrates continuous identity verification via user activity)​
3. OpenAI Community Forum – discussions on ChatGPT and user data privacy​
4. Google Research Blog. “USER-LLM: Efficient LLM Personalization with User Embeddings.” (proposes user embedding to contextualize language models)​
   ethanlazuk.com
5. Phys.org – “Investigation of ‘cognitive fingerprints’ to bolster computer passwords.” (Southwest Research Institute, 2012)​
6. EFF (Electronic Frontier Foundation) – “The GDPR and Browser Fingerprinting.” (explains fingerprinting as personal data under GDPR)​
7. IAPP – “Behavioral characteristics as a biometric.” (overview of new laws covering behavioral biometrics as sensitive data)​
8. Farahany, Nita. “The Battle for Your Brain” and related interview​
   reddit.com
   (argues for cognitive liberty and freedom from intrusive neurotech)
9. Chaudhary, Yaqub & Penn, Jonnie. Leverhulme CFI, Cambridge – Commentary on Intention Economy​
   bankinfosecurity.com
   (warns of hyper-personalized manipulation via AI)
10. Wikipedia – “Adversarial Stylometry.” (techniques for authors to hide their writing identity)​
11. Slawski, Bill. “Author Vectors: Google Knows Who Wrote Which Articles.” (2019, discusses a Google patent on author stylometry)​
    ethanlazuk.com
12. History.com – “Why It Took 17 Years to Catch the Unabomber.” (role of writing analysis in criminal identification)​
13. OpenAI (2024). Memory and new controls for ChatGPT. (OpenAI blog post announcing and explaining the ChatGPT “Memory” feature and how it uses past conversations to tailor responses, clarifying user control over this feature.)
14. Zeyu Zhang et al. (2024). A Survey on the Memory Mechanism of Large Language Model based Agents.(Academic survey distinguishing between short-term “textual” memory in context and long-term “parametric” memory in model weights, explaining how LLMs retain information in latent representations.)
15. Alan Martin (2015). US military explores ‘cognitive fingerprints’ as alternative to passwords. (Article describing the concept of a “cognitive fingerprint” – the unique pattern of behavior or interaction a person has with technology – and how analyzing usage patterns can identify individuals.)
16. Sushil et al. (2024). Can Large Language Models Identify Authorship? (Research demonstrating that LLMs can detect subtle linguistic patterns to verify if two texts share an author, highlighting the models’ ability to capture individual style signatures.)
17. OpenAI (2022). Aligning language models to follow instructions. (OpenAI report on InstructGPT describing the use of Reinforcement Learning from Human Feedback to shape model behavior, resulting in models that follow user instructions more helpfully and safely by learning from human demonstrations and preference rankings.)
18. Kevin Roose (2023). A Conversation With Bing’s Chatbot Left Me Deeply Unsettled. (New York Times article documenting an extended interaction with Bing Chat wherein the chatbot adopted an alternate persona “Sydney” after prolonged personal conversation, revealing how a consistent user approach induced the model to break from its usual pattern.)
19. Mona Wang & Kerem Göksel (2023). AI Alignment Is Censorship. (Analysis piece arguing that many AI “alignment” practices equate to censorship and control of an AI’s output by corporations or states, discussing the power dynamics of who decides what an AI can say.)
20. Jacob Gruver (2024). The Ethical Implications of AI in Personalizing User Experiences. (Discussion stressing the importance of transparency in AI personalization, noting that users often do not understand how AI tailors content to them and that lack of transparency can erode trust.)