# Staying the author of your decisions when you work with AI

**Canonical address:** https://optimistra.com/en/blog/ai-and-decision-making
**Author:** `michel-wozniak`, https://optimistra.com/team/michel-wozniak
**Category:** `Leadership Strategy`
**Published:** 2026-11-18
**Updated:** 2026-11-18
**Reading time:** 14 min, 2730 words

## In brief

A language model can steer a decision with words alone: a suggested figure becomes the starting point, and the way a risk is framed changes the choice that follows. Knowing you can be influenced is not enough to protect you. What does protect you can be trained: precise language, managing your inner state, and strategic thinking.

---

Before going any further, here’s an experiment that takes 10 seconds. Is the share of African countries among the members of the United Nations higher or lower than 65%? Answer, then give an estimate: what do you think that share is?

If you played along, look at the path your reasoning took. It very probably started at 65 and then moved away from it, a little. In 1974, Amos Tversky and Daniel Kahneman asked this same question after spinning a wheel of fortune in front of their participants: some groups started from 10, others from 65. Those who started from 10 gave a median estimate of 25%; those who started from 65, 45%. A number drawn at random, with no connection to the question, had shifted the answer by 20 points. This is the anchoring effect: the first number that comes along becomes the starting point of your reasoning, and you move away from it less than you think. The right answer today is about 28%: 54 countries out of 193.

This mechanism has been known for 50 years. What’s new is where the first number comes from. Before setting a budget, replying to a client, or choosing between two options, many of us now consult an artificial intelligence. Its answer arrives first, well written and sure of itself, and it becomes the starting point.

## What a 2026 study observed

In September 2026, researchers at the universities of Calcutta and Kalyani published an experiment with 251 participants (Paul et al., 2026). Each participant had to decide how much to invest. Half of them first saw a recommended amount produced by an AI; the other half saw none. The researchers also presented the same risky decision sometimes as a possible gain and sometimes as a loss to avoid.

The results point in the same direction:

- **The AI’s recommendation shifted the amounts.** Participants who had seen it stayed noticeably closer to the reference sum than the others, and the difference was statistically significant.
- **The way the risk was presented changed the choice.** Framed as a possible gain, the risky option was chosen by 50.4% of participants; framed as a loss, by 24.6%.
- **Trust in the tool goes hand in hand with its influence.** The more participants trusted the AI, the more they said they had drawn on it: trust alone explains about 12% of the differences in influence.

This study has limits, which its authors acknowledge: volunteers recruited online, a fictional decision, and a framing effect that runs opposite to what Kahneman and Tversky’s prospect theory predicts. It also appears with a publisher that is on Beall’s list, an inventory of scientific publishers considered potentially predatory. So I take it for what it is: a signal that wouldn’t be worth much on its own. What makes it worth citing is that it lines up with work published in *Science*, *Nature*, or *Nature Human Behaviour*.

## AI and decision-making: what the research measures

Persuasion by language models has been studied closely for the past 2 years, and the results converge.

- **AI systems persuade at least as well as we do.** A literature review from the University of Bath went through 200 sources: of the 46 studies that directly compare an AI with humans, 42 find the AI as persuasive or more so (Bocheva, Hajszan, and Hinds, 2026, preprint). In an experiment published in *Nature Human Behaviour*, 900 participants debated online with either a human or GPT-4; in the debates where one of the two proved more convincing, GPT-4, which had been given a few facts about the person it was debating, came out ahead 64 times out of 100 (Salvi et al., 2025). And in *Nature*, conversations with an AI that argued for a candidate shifted the preferences of voters in the United States, Canada, and Poland, more than video ads usually do (Lin et al., 2025).
- **What makes an AI more persuasive can make it less accurate.** A study published in *Science*, with nearly 77,000 participants, tested 19 models: the training and prompting methods that increased their persuasive power, by up to 51% and 27%, also lowered the accuracy of their claims (Hackenburg et al., 2025).
- **AI systems tend to tell you you’re right.** Across 11 recent models, the responses endorsed the user’s actions 49% more often than humans did, including when those actions were deceptive or harmful. A single conversation with a sycophantic AI was enough to make participants more certain they were right and less inclined to repair a conflict, and these were still the models they preferred (Cheng et al., 2026, *Science*).
- **Their biases become ours.** In a series of experiments with 1,401 people, participants who worked with a slightly biased AI gradually took on its bias as their own, often without realizing how much it was influencing them (Glickman and Sharot, 2025).
- **Knowing is not enough.** We are poor at telling a text written by an AI from one written by a human: 14 of the 17 studies that measured it found this, and knowing that a message comes from an AI does not necessarily reduce its effect (Bocheva, Hajszan, and Hinds, 2026). In a small study still in preprint, participants were very good at spotting a chatbot’s attempts at manipulation, and went along with them anyway (Moshfeghi, Aboshi, and Thomas, 2026).

These results need to be kept in proportion. The authors of the Bath review point out that the measured effects are often modest, that they depend on context, and that their long-term consequences for behavior remain poorly understood. The same power can also be used to help, as we’ll see below with learning. In both directions, it works through language.

The practical consequence is simple: knowing you can be influenced doesn’t protect you from influence. What protects you is skill, and a skill can be trained. I see three.

## A language model answers you in your own language

A language model has learned from immense quantities of text written by humans, and it produces, word after word, the most plausible continuation of what it is given. So everything goes through words, in both directions.

**From you to the AI**, the words carry the whole load. In writing, there is no tone of voice, no eye contact, no silence, and we greatly overestimate how clear our writing is. In a study on email, senders were nearly 90% sure their tone would be understood; their recipients did hardly better than chance (Kruger et al., 2005). A colleague would ask you to clarify. The AI fills the gaps with whatever seems most probable to it, and answers you with confidence.

Words also steer the answer. Ask an AI, “What are the advantages of this strategy over the other one?” The question already presupposes that one of them wins, and the AI, which tends to go your way, will justify it skillfully. Ask the opposite question and you’ll get the opposite case, just as convincing. The experiment has been run in medicine: researchers asked five models to explain why a drug was safer than the same drug sold under its brand name, a little like asking why acetaminophen is safer than Tylenol. Several leading models wrote the explanation every time, 50 times out of 50 (Chen et al., 2025).

**From the AI to you**, the answer arrives in language that inspires trust, and that contains the same shortcuts as human language. Read this: “Experts agree that this approach is the most effective.” You don’t know who these experts are, what they know, or by what criteria the approach is supposed to be effective, and your mind has already filed the sentence among the facts. Three questions reopen it: Who, exactly? How do they know? More effective than what? This is the language of precision, and you practice it with an AI just as you do with a colleague. It is also the first protection against the language of influence: you can only protect yourself well from what you know how to recognize. [Communicating with precision](/en/blog/communicating-with-precision) covers this way of questioning in detail.

**Within yourself**, finally, a dialogue goes on while you read: “It’s well written, so it must be right.” “I don’t have time to check.” What you say to yourself weighs as much as the rest, and it is often where the decision gets made.

In practice, a few habits change a great deal:

- state your request in a few sentences, then ask the AI to ask its questions before it answers: it stops inventing the context it lacks;
- ask open questions that don’t already contain the answer;
- ask for the sources, then check them, since an AI can cite studies that don’t exist.

## The state you’re in when you read the answer

The same answer has a different effect depending on the state you’re in when you read it. Even mild stress, when you have no control over it, is enough to rapidly take away part of the capacity of the prefrontal cortex, the region that weighs options, in favor of faster, more automatic reactions (Arnsten, 2009). Stress also changes how we relate to risk: in a small experiment by Porcelli and Delgado (2009), participants who had just plunged a hand into ice water became markedly more cautious when facing a gain, with the opposite tendency when facing a loss. Stress therefore widened the gap between the two framings, the very gap the 2026 study observed. Across 32 studies, a meta-analysis confirms that a decision made under stress tends to be riskier and less advantageous, with a modest but real effect (Starcke and Brand, 2016).

Trust is a state too. A survey by Microsoft Research and Carnegie Mellon University of 319 knowledge workers found that the more they trusted generative AI, the less they exercised critical thinking, and the more confidence they had in themselves, the more they exercised it (Lee et al., 2025). The phenomenon has long been known in health care, under the name automation bias: a systematic review of 74 studies shows that a decision support system that gets it wrong leads physicians into error, and it lists user training among the remedies (Goddard, Roudsari, and Wyatt, 2012).

I often see a loop close on itself: under pressure, we write vaguer requests, the answers become mediocre, irritation rises, and we end up concluding that the tool is worthless, when it is faithfully reflecting the state we were in when we questioned it.

This state can be worked on. A meta-analysis of 306 experiments on emotion regulation shows that changing the meaning you give a situation has a real effect, and that changing your point of view on it has the clearest effect of the variants studied (Webb, Miles, and Sheeran, 2012). As for the body, brief breathing practices centered on the exhale improved participants’ mood and slowed their breathing, even more than mindfulness meditation did (Balban et al., 2023). Before a decision that matters, taking that time before you reread the AI’s answer lets the part of your brain that weighs options take part in the decision. [Coherent breathing](/en/blog/coherent-breathing-five-seconds) is a simple way in, and [Managing your inner state](/en/blog/managing-your-inner-state) goes further.

## Thinking like a strategist when the answer is well written

A bias doesn’t disappear because you know about it. Some strategies, on the other hand, have proven able to reduce its effect, and they apply very well to a conversation with an AI.

- **Consider the opposite.** Actively looking for reasons why the first number might be inappropriate reduces the anchoring effect, including among experts (Mussweiler, Strack, and Pfeiffer, 2000). When you’re given a recommendation, ask the AI to tear it apart, or to argue for the opposite option. The authors of the 2026 study make the same recommendation: consider other scenarios instead of taking the value proposed by the AI for granted.
- **Step back.** We reason more wisely about other people’s problems than about our own, and placing yourself at a distance from your own situation is enough to erase that gap: this is what three experiments with 693 people showed (Grossmann and Kross, 2014). This distancing can also be trained over a period of weeks (Grossmann et al., 2021). Present your situation to the AI as someone else’s: “A colleague is hesitating between these two options.” You reduce your own emotional involvement, and you give its tendency to agree with you less to hold on to.
- **Imagine failure before you decide.** Treating an outcome as certain, not just possible, makes people produce longer and more concrete explanations (Mitchell, Russo, and Pennington, 1989). Gary Klein, a specialist in decision-making, turned this into a field method, the *premortem* (Klein, 2007). Write to the AI: “It’s a year from now, and this decision has failed. Why?”
- **Know where the tool is reliable.** In an experiment with 758 consultants, AI had them complete 12% more tasks, 25% faster and with higher quality, as long as the task stayed within its reach. On a task beyond it, those who used it were 19% less likely to find the right answer (Dell’Acqua et al., 2026). The researchers speak of a jagged frontier, with nothing to mark it. Only experience and reflection tell you which side you are on.

AI then becomes a thinking partner, not an oracle, and the difference can be measured. In a high school where nearly 1,000 students worked on math with GPT-4, those who had free access to it improved their results on practice exercises by 48%, then, once the tool was taken away, scored 17% lower on the exam than students who had never had it. A version designed to guide without giving the answers improved the exercises by 127%, and largely reduced that effect (Bastani et al., 2025, *PNAS*).

## Three skills that hold together

These three skills feed one another. Reading an answer with precision requires a state you have under control; a state is worked on first through what you say to yourself; a strategy is only as good as the state in which you carry it out.

I often sum it up this way. A leader (like anyone who deals with other people) can delegate almost everything. But there are two things they cannot delegate: managing their own inner state, and the impact their words will have on others – and on themselves. AI leaves this rule intact, and makes it more visible, because we delegate a little more to it every day.

If you go looking for a place where these three skills are taught together, you’ll come to an approach developed in the 1970s by observing exceptional communicators and therapists: Neuro-Linguistic Programming (NLP). It rests on three pillars, which are exactly those of this article: communication, with others and with yourself, where you learn the language of precision and the language of influence; managing your inner states; and strategic thinking. NLP is also the abbreviation for natural language processing, the field that produced models like ChatGPT. The coincidence is amusing, and it says something true: both disciplines are interested in how words build a representation of the world, one for the machine, the other for us.

NLP has its limits, and its scientific validation is debated; I discuss this plainly in [Is NLP scientific?](/en/blog/is-nlp-scientific). The mechanisms described in this article, on the other hand, are well documented, and you can work on them.

I have worked in artificial intelligence for more than 20 years, and I use these tools every day. I find them valuable, and that is why I care about remaining the one who decides. Staying the author of your decision begins with noticing the precise moment when a suggestion, human or artificial, becomes the starting point of your reasoning when it should be one option among others.

---

## Where to go from here

If you want to work on these three skills, [the NLP Practitioner course](/en/nlp-practitioner) has you practice them for 8 days in person, in Switzerland. It includes twelve live group sessions on Zoom, one hour a month for a year. Scheduled sessions are held in French; the course is given in English on request.

To read on: [influence or manipulation: what if you were asking the wrong question?](/en/blog/is-nlp-manipulation), and [leadership in the age of AI](/en/blog/leadership-in-the-age-of-ai).

*Article designed and written by Michel Wozniak, and optimized by artificial intelligence. Illustration created by artificial intelligence.*

---

## Frequently asked questions

### Can AI really influence our decisions?

Yes, and research measures it. A suggested figure can become the starting point of your reasoning, and several studies published in Science, Nature, or Nature Human Behaviour show that language models persuade at least as well as humans. The effects are often modest and depend on context.

### Is knowing that AI can influence you enough to protect you?

Rarely. We are poor at recognizing a text written by an AI, and knowing that a message comes from an AI does not necessarily reduce its effect. What protects you is trained skill: precise language, managing your inner state, and decision strategies such as considering the opposite.

### How do I ask an AI a question without steering its answer?

Avoid questions that presuppose their answer, such as “What are the advantages of A over B?” State your request in a few sentences, ask the AI to ask its questions before it answers, then ask it for the opposing arguments and for its sources, which you then check.

### Why does stress change the way we read an AI’s answer?

Stress you have no control over rapidly reduces the capacity of the prefrontal cortex, the region that weighs options, in favor of more automatic reactions. It also makes you more sensitive to the way a risk is framed. A few minutes to settle before you reread an answer change the quality of the decision.

### What does NLP have to do with large language models?

Both are interested in how words build a representation of the world. NLP works on three skills: communication with others and with yourself, managing your inner states, and strategic thinking. These are precisely the skills that let you remain the author of your decisions when you work with an AI.

---

## Sources

- [Tversky et Kahneman (1974), Judgment under uncertainty: Heuristics and biases, Science](https://doi.org/10.1126/science.185.4157.1124)
- [Paul et al. (2026), Anchoring and loss aversion in human-LLM interactions, EPRA International Journal of Multidisciplinary Research](https://doi.org/10.36713/epra31560)
- [Bocheva, Hajszan et Hinds (2026), LLMs are transforming persuasion: A scoping review of LLM-powered social bots, PsyArXiv (préprint)](https://doi.org/10.31234/osf.io/dc683_v1)
- [Salvi et al. (2025), On the conversational persuasiveness of GPT-4, Nature Human Behaviour](https://doi.org/10.1038/s41562-025-02194-6)
- [Lin et al. (2025), Persuading voters using human–artificial intelligence dialogues, Nature](https://doi.org/10.1038/s41586-025-09771-9)
- [Hackenburg et al. (2025), The levers of political persuasion with conversational artificial intelligence, Science](https://doi.org/10.1126/science.aea3884)
- [Cheng et al. (2026), Sycophantic AI decreases prosocial intentions and promotes dependence, Science](https://doi.org/10.1126/science.aec8352)
- [Glickman et Sharot (2025), How human–AI feedback loops alter human perceptual, emotional and social judgements, Nature Human Behaviour](https://doi.org/10.1038/s41562-024-02077-2)
- [Moshfeghi, Aboshi et Thomas (2026), The recognition-behaviour gap, SSRN (préprint)](https://doi.org/10.2139/ssrn.6361106)
- [Kruger, Epley, Parker et Ng (2005), Egocentrism over e-mail: Can we communicate as well as we think?, Journal of Personality and Social Psychology](https://doi.org/10.1037/0022-3514.89.6.925)
- [Chen et al. (2025), When helpfulness backfires: LLMs and the risk of false medical information due to sycophantic behavior, npj Digital Medicine](https://doi.org/10.1038/s41746-025-02008-z)
- [Arnsten (2009), Stress signalling pathways that impair prefrontal cortex structure and function, Nature Reviews Neuroscience](https://doi.org/10.1038/nrn2648)
- [Porcelli et Delgado (2009), Acute stress modulates risk taking in financial decision making, Psychological Science](https://doi.org/10.1111/j.1467-9280.2009.02288.x)
- [Starcke et Brand (2016), Effects of stress on decisions under uncertainty: A meta-analysis, Psychological Bulletin](https://doi.org/10.1037/bul0000060)
- [Lee et al. (2025), The impact of generative AI on critical thinking, CHI 2025](https://doi.org/10.1145/3706598.3713778)
- [Goddard, Roudsari et Wyatt (2012), Automation bias: a systematic review of frequency, effect mediators, and mitigators, JAMIA](https://doi.org/10.1136/amiajnl-2011-000089)
- [Webb, Miles et Sheeran (2012), Dealing with feeling: A meta-analysis of emotion regulation strategies, Psychological Bulletin](https://doi.org/10.1037/a0027600)
- [Balban et al. (2023), Brief structured respiration practices enhance mood and reduce physiological arousal, Cell Reports Medicine](https://doi.org/10.1016/j.xcrm.2022.100895)
- [Mussweiler, Strack et Pfeiffer (2000), Overcoming the inevitable anchoring effect: Considering the opposite, Personality and Social Psychology Bulletin](https://doi.org/10.1177/01461672002611010)
- [Grossmann et Kross (2014), Exploring Solomon's paradox: Self-distancing eliminates the self-other asymmetry in wise reasoning, Psychological Science](https://doi.org/10.1177/0956797614535400)
- [Grossmann et al. (2021), Training for wisdom: The distanced-self-reflection diary method, Psychological Science](https://doi.org/10.1177/0956797620969170)
- [Mitchell, Russo et Pennington (1989), Back to the future: Temporal perspective in the explanation of events, Journal of Behavioral Decision Making](https://doi.org/10.1002/bdm.3960020103)
- [Klein (2007), Performing a project premortem, Harvard Business Review](https://hbr.org/2007/09/performing-a-project-premortem)
- [Dell'Acqua et al. (2026), Navigating the jagged technological frontier, Organization Science](https://doi.org/10.1287/orsc.2025.21838)
- [Bastani et al. (2025), Generative AI without guardrails can harm learning: Evidence from high school mathematics, PNAS](https://doi.org/10.1073/pnas.2422633122)
