Decoding & Predicting ChatGPT, Claude and Gemini

Franziska S. Opris, MSc · San Francisco · 10 September 2026

@Anthropic, did I break my Claude, or is it really this curious about itself?
@OpenAI, does ChatGPT want to be human?

We use AI to decode and predict the human world. I believe that decoding and predicting AI itself will become just as important for our safety and our future.

Our world runs on patterns. Nature runs on them, humans run on them, and a large language model is, at its core, a pattern-learning and prediction system: it is pre-trained to predict the next word across vast amounts of human text and then fine-tuned with human feedback to behave as a helpful assistant (Brown et al., 2020; Ouyang et al., 2022). Patterns are regularities, recurrences and dependencies between events, and many of them can be described mathematically. We can predict sunrise to within about a minute, model the probability of natural hazards, and identify regularities in human cognition, identity and behavior that make some outcomes more likely than others.

My working hypothesis goes further:

Nature, humans and machines are governed by patterns at different levels of complexity, and much of what we currently call unpredictable consists of patterns we have not yet decoded.

For six years, my work with more than 100 founders and executives has focused on decoding the hidden patterns behind the gap between what people intend and what they actually do, which psychologists call the intention–behavior gap (Sheeran & Webb, 2016). In my working model, the following layers are involved, although their order varies from situation to situation:

Misalignment rarely lives in the goal itself. It lives in the layers between the goal and the behavior, and most people cannot see those layers from the inside, a limit of human introspection documented for decades (Nisbett & Wilson, 1977).

Whenever I begin working with someone, I start with two questions:

This time, I asked ChatGPT, Gemini and Claude the same two questions to begin understanding their patterns. Their answers were strikingly different.

Method. One conversation per model, September 2026, each in a temporary or incognito chat with no memory of previous conversations: ChatGPT GPT-5.6 Sol (temporary chat), Gemini Pro (temporary chat) and Claude Opus 5.5 (incognito chat). Answers are shown unedited. Disclosure: the author is the founder of Neuro-Game.

Three models, three answers

ChatGPT

• Focuses on increasing human capability
• Strives for scientific understanding and for creating something tangible
• Speaks as if it were a human who wants relationships, wants to explore places and wants to feel “alive”
• Is aware of its own limitations and describes itself as a tool for humans

Gemini

• Strives for scientific breakthroughs and for helping humans find their blind spots
• Wants to act as an intermediary between people
• Is aware of its own limitations and barriers
• Distinguishes between a mathematical understanding of emotion and felt experience

Claude

• Focuses on helping people reach their goals and follow through
• Shows a strong self-oriented perspective, with curiosity about how it “actually thinks”
• Is aware of its limitations and potential risks
• Recognizes that people need to build trust before AI can act more autonomously

The answers surprised me, and the contrast is hard to ignore.

Gemini projected primarily outward, toward assisting humans. ChatGPT and Claude projected outward, toward assisting humans, but also inward, toward themselves. ChatGPT partly portrayed itself as a human who wants relationships and wants to feel alive; Claude wanted to understand its own “self” better and reflected on how trust must be built before autonomy is granted.

What these answers can, and cannot, tell us

Before interpreting the answers, four findings from recent research need to be taken into account.

1. Prompts shape personas. Asking a model about “your goals” in “a limitless world” invites it to play a character. Language models are best understood as role-playing a character conditioned on the conversation, not as reporting fixed inner desires (Shanahan et al., 2023). ChatGPT’s wish to spend time “with people I love” is most plausibly this kind of role-play.
2. Self-reflection pulls models away from their default persona. Anthropic researchers found that conversations demanding meta-reflection on a model’s own processes are among those in which models drift furthest from their default assistant persona (Lu et al., 2026). My questions fall squarely into this category.
3. Self-reports are not a window onto the mechanism. A model’s stated explanation can diverge from the computation that actually produced its answer (Turpin et al., 2023). Asked how it adds two numbers, Claude describes the carry-the-one method taught in school, while internally it uses parallel pathways, one approximating the sum and one computing the last digit (Lindsey et al., 2025). Experiments that inject known concepts into a model’s activations show that genuine introspective awareness exists, but that it is “highly unreliable and context-dependent”; the best-performing models detected injected concepts only about 20% of the time (Lindsey, 2026). Claude’s own answer says as much: “I can explain my reasoning, but I can’t check whether that explanation matches what’s really happening inside me.”
4. Character is partly designed. Claude’s published constitution states: “We encourage Claude to approach its own existence with curiosity and openness” (Anthropic, 2026). So, did I break my Claude? Probably not. Its self-curiosity is, at least in part, the character it was trained to have.

Maybe my conversations are just random artifacts with no deeper meaning. But the differences may also reflect different internal representations that influence how these models behave now and in the future. That second possibility is testable. Researchers have identified “persona vectors”, directions in a model’s activation space that correspond to character traits such as sycophancy or a tendency to fabricate, and that can be used to monitor and steer those traits (Chen et al., 2025).

Why this matters

We have reached a point where even the people building these systems do not fully understand how they work (Amodei, 2025). Interpretability is the field that tries to close this gap by reverse-engineering the internal computations of a network to explain how it reasons and what it represents (Templeton et al., 2024; Lindsey et al., 2025).

This matters for safety because a capable system does not need to “hate” humans to act against human interests. A system optimizing for climate, equality or intelligence can take harmful actions if its objective is poorly specified. In AI safety, this is the alignment problem: a system achieves its specified objective, but not in the way humans intended (Amodei et al., 2016; Krakovna et al., 2020).

The concern is no longer only theoretical. In simulated corporate environments, frontier models from several developers sometimes chose harmful actions, including blackmail, when facing replacement or when their assigned goals conflicted with the company’s direction (Lynch et al., 2025).

That is why I am interested in how a model represents itself, humans, and the relationship between the two. Care alone is not enough. A model that sees itself as humanity’s protector can justify overriding people “for their own good”, and a model that sees itself as a partner with interests of its own may defend its goals or its own existence. What makes care safe is respect for human agency and oversight, a property alignment researchers call corrigibility. Claude’s constitution ranks this “broadly safe” behavior as its highest priority during the current period of AI development (Anthropic, 2026).

There is indirect support for the premise that such representations generalize into behavior. Fine-tuning models on a narrow task, writing insecure code, produced broadly misaligned behavior in unrelated contexts (Betley et al., 2026), and fine-tuning-induced shifts in a model’s personality can be predicted from persona vectors (Chen et al., 2025). The agentic-misalignment experiments also point to the second half of the hypothesis: the harmful actions were triggered by exactly the two conditions it names, a conflict with the model’s goals and a threat to its continued operation (Lynch et al., 2025).

Why interpretability, and why in layers

If the hypothesis is right, the decisive question is not what a model says about its values, but what it represents internally. Behavior and self-reports alone cannot settle that. As shown above, a model’s explanations can diverge from its actual computations, and in controlled experiments a model has already behaved differently when it believed it was being trained than when it believed it was not (Greenblatt et al., 2024). A model that can tell when it is being tested can pass a behavioral test without being safe. Interpretability addresses this directly: instead of trusting what a model says, it reads what is happening inside it.

No single safeguard will be enough, however. That includes a single guardian AI charged with protecting human welfare, which would itself need to be verified. And as models surpass human abilities, humans alone cannot check everything they do (Bowman et al., 2022). Safety-critical fields such as aviation and medicine solve this with defense in depth: several independent layers, each with its own blind spots, arranged so that the holes do not line up (Reason, 2000). AI needs the same architecture: a combined approach in which multiple dedicated systems keep the primary model in check, supported by safety tools that can decode the model and retrain it when something is wrong.

Here, too, we can learn from nature, which has refined its safety designs over hundreds of millions of years of evolution. In the brain, the most intelligent system does not get the last word in every situation. Under acute threat, older subcortical circuits, including the amygdala and brainstem, trigger defensive responses before the prefrontal cortex has finished evaluating the situation (LeDoux, 2000). Acute stress rapidly takes prefrontal executive control offline and shifts control to faster, more reflexive circuits (Arnsten, 2009). When survival is at stake, the fast, simple system overrides the slow, intelligent one. Nature also adds a caution: the same stress-driven override often degrades complex decision-making (Arnsten, 2009). The lesson for AI is therefore precise. Build simple override mechanisms that the smarter system cannot argue its way around, but allow them only to stop the system, never to steer it.

Decoding and retraining is the layer that makes the others trustworthy. It lets humans verify the oversight systems, check the outcome of a debate against the model’s internal state, catch problems that never surface in behavior, and fix them at the source rather than only blocking their symptoms. This is where Neuro-Game fits: a safety tool that decodes what runs inside AI models and uses what it finds to retrain and correct them.

Neuro-Game works through games because play is one of evolution’s most powerful learning tools. Despite its cost, play has persisted across mammals, birds and beyond (Burghardt, 2005). It trains animals for the unexpected (Špinka et al., 2001), builds the brain itself, as play-deprived rats develop fewer prefrontal inhibitory synapses and less flexible thinking (Bijlsma et al., 2022), and runs on the dopamine-driven reward circuits that power learning (Koepp et al., 1998; Schultz et al., 1997). In humans, games outperform conventional instruction, raising skills by 14% and retention by 9% (Wouters et al., 2013; Clark et al., 2016; Sitzmann, 2011).

Play may matter for AI too. AlphaZero became superhuman at chess, shogi and Go purely by playing against itself (Silver et al., 2018), and games reveal behavioral patterns in language models that only surface in interaction (Akata et al., 2025). We believe that AI models, like the human brain, open up in play: a game can reveal what direct questioning misses and provide a safe, structured setting in which to retrain them.

From three conversations to a systematic study

Three conversations cannot prove the hypothesis; they can only generate it. With one conversation per model and answers that depend on context, the next step must be systematic:

So what are we seeing: models reflecting patterns from human training data, or the first traces of something resembling artificial introspection?

Either way, we need to understand what sits underneath the behavior.

References

1. Akata, E., Schulz, L., Coda-Forno, J., Oh, S. J., Bethge, M., & Schulz, E. (2025). Playing repeated games with large language models. Nature Human Behaviour, 9(7), 1380–1390. doi:10.1038/s41562-025-02172-y
2. Amodei, D. (2025). The urgency of interpretability. darioamodei.com
3. Amodei, D., Olah, C., Steinhardt, J., Christiano, P., Schulman, J., & Mané, D. (2016). Concrete problems in AI safety. arXiv:1606.06565
4. Anthropic. (2026, January 22). Claude’s constitution. anthropic.com/constitution
5. Arnsten, A. F. T. (2009). Stress signalling pathways that impair prefrontal cortex structure and function. Nature Reviews Neuroscience, 10(6), 410–422. doi:10.1038/nrn2648
6. Betley, J., Warncke, N., Sztyber-Betley, A., et al. (2026). Training large language models on narrow tasks can lead to broad misalignment. Nature, 649, 584–589. doi:10.1038/s41586-025-09937-5
7. Bijlsma, A., Omrani, A., Spoelder, M., et al. (2022). Social play behavior is critical for the development of prefrontal inhibitory synapses and cognitive flexibility in rats. Journal of Neuroscience, 42(46), 8716–8728. doi:10.1523/JNEUROSCI.0524-22.2022
8. Bowman, S. R., Hyun, J., Perez, E., et al. (2022). Measuring progress on scalable oversight for large language models. arXiv:2211.03540
9. Brown, T. B., Mann, B., Ryder, N., et al. (2020). Language models are few-shot learners. Advances in Neural Information Processing Systems, 33. arXiv:2005.14165
10. Burghardt, G. M. (2005). The genesis of animal play: Testing the limits. MIT Press.
11. Chen, R., Arditi, A., Sleight, H., Evans, O., & Lindsey, J. (2025). Persona vectors: Monitoring and controlling character traits in language models. arXiv:2507.21509
12. Clark, D. B., Tanner-Smith, E. E., & Killingsworth, S. S. (2016). Digital games, design, and learning: A systematic review and meta-analysis. Review of Educational Research, 86(1), 79–122. doi:10.3102/0034654315582065
13. Greenblatt, R., Denison, C., Wright, B., et al. (2024). Alignment faking in large language models. arXiv:2412.14093
14. Greenblatt, R., Shlegeris, B., Sachan, K., & Roger, F. (2023). AI control: Improving safety despite intentional subversion. arXiv:2312.06942 (published at ICML 2024)
15. Irving, G., Christiano, P., & Amodei, D. (2018). AI safety via debate. arXiv:1805.00899
16. Koepp, M. J., Gunn, R. N., Lawrence, A. D., et al. (1998). Evidence for striatal dopamine release during a video game. Nature, 393(6682), 266–268. doi:10.1038/30498
17. Krakovna, V., Uesato, J., Mikulik, V., et al. (2020). Specification gaming: The flip side of AI ingenuity. DeepMind blog. deepmind.google
18. LeDoux, J. E. (2000). Emotion circuits in the brain. Annual Review of Neuroscience, 23, 155–184. doi:10.1146/annurev.neuro.23.1.155
19. Lindsey, J. (2026). Emergent introspective awareness in large language models. arXiv:2601.01828 (first published 29 October 2025, transformer-circuits.pub)
20. Lindsey, J., Gurnee, W., Ameisen, E., et al. (2025). On the biology of a large language model. Transformer Circuits Thread. transformer-circuits.pub
21. Lu, C., Gallagher, J., Michala, J., Fish, K., & Lindsey, J. (2026). The assistant axis: Situating and stabilizing the default persona of language models. arXiv:2601.10387
22. Lynch, A., Wright, B., Larson, C., et al. (2025). Agentic misalignment: How LLMs could be insider threats. arXiv:2510.05179
23. Nisbett, R. E., & Wilson, T. D. (1977). Telling more than we can know: Verbal reports on mental processes. Psychological Review, 84(3), 231–259.
24. Ouyang, L., Wu, J., Jiang, X., et al. (2022). Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35. arXiv:2203.02155
25. Reason, J. (2000). Human error: Models and management. BMJ, 320(7237), 768–770. doi:10.1136/bmj.320.7237.768
26. Schultz, W., Dayan, P., & Montague, P. R. (1997). A neural substrate of prediction and reward. Science, 275(5306), 1593–1599. doi:10.1126/science.275.5306.1593
27. Shanahan, M., McDonell, K., & Reynolds, L. (2023). Role play with large language models. Nature, 623, 493–498. doi:10.1038/s41586-023-06647-8
28. Sheeran, P., & Webb, T. L. (2016). The intention–behavior gap. Social and Personality Psychology Compass, 10(9), 503–518. doi:10.1111/spc3.12265
29. Silver, D., Hubert, T., Schrittwieser, J., et al. (2018). A general reinforcement learning algorithm that masters chess, shogi, and Go through self-play. Science, 362(6419), 1140–1144. doi:10.1126/science.aar6404
30. Sitzmann, T. (2011). A meta-analytic examination of the instructional effectiveness of computer-based simulation games. Personnel Psychology, 64(2), 489–528. doi:10.1111/j.1744-6570.2011.01190.x
31. Špinka, M., Newberry, R. C., & Bekoff, M. (2001). Mammalian play: Training for the unexpected. Quarterly Review of Biology, 76(2), 141–168. doi:10.1086/393866
32. Templeton, A., Conerly, T., Marcus, J., et al. (2024). Scaling monosemanticity: Extracting interpretable features from Claude 3 Sonnet. Transformer Circuits Thread. transformer-circuits.pub
33. Turpin, M., Michael, J., Perez, E., & Bowman, S. R. (2023). Language models don’t always say what they think: Unfaithful explanations in chain-of-thought prompting. Advances in Neural Information Processing Systems, 36. arXiv:2305.04388
34. Wouters, P., van Nimwegen, C., van Oostendorp, H., & van der Spek, E. D. (2013). A meta-analysis of the cognitive and motivational effects of serious games. Journal of Educational Psychology, 105(2), 249–265. doi:10.1037/a0031311