Theory of Mind & AI: The Sally-Anne Test Explains Why Your Prompts Fail

I asked AI to summarize a 40-page vendor contract last week. It returned four paragraphs of the most confident, most useless boilerplate I’ve ever read. The pricing model was wrong. The termination clause was missing. And it invented a 90-day notice requirement that wasn’t in the document.

The AI wasn’t broken. I just didn’t tell it what mattered.

This happens constantly. You ask AI something specific, you get something generic. You feed it real data, you get surface-level observations you already knew. The frustration is universal, and it has nothing to do with which model you use or how much you pay. It’s a theory of mind problem. And AI doesn’t have one.

There’s a name for this skill gap, and it isn’t prompt engineering. But we’ll get to that.

The real explanation starts with two dolls, a marble, and a psychology experiment from 1985. This post is adapted from a presentation I delivered at the 2026 HFTP National Club Summit in Austin.

The Sally-Anne Test: A False Belief Task

In 1985, three researchers at the MRC Cognitive Development Unit in London ran an experiment on children. Simon Baron-Cohen, Alan Leslie, and Uta Frith wanted to test something deceptively simple: can a child understand that another person can believe something that isn’t true?

Here’s how it worked.

A child watches two dolls act out a short scene. Sally has a basket. Anne has a box. Sally puts a marble in her basket, then leaves the room. While Sally is gone, Anne takes the marble out of Sally’s basket and puts it in her own box.

Sally comes back. The researcher asks the child: “Where will Sally look for her marble?”

Think about that for a second. The marble is in the box. But Sally doesn’t know that. Sally still believes the marble is where she left it, in the basket.

To answer correctly, the child has to do something remarkable. They have to set aside what they know (the marble is in the box) and model what Sally knows (she thinks it’s still in the basket). They have to hold two realities in their head simultaneously: the world as it actually is, and the world as someone else believes it to be.

85% of neurotypical children over age four got this right. They pointed to the basket. They understood that Sally held a false belief, a belief that didn’t match reality, and that she would act on her belief, not on the truth.

80% of the autistic children in the study pointed to the box. They answered based on where the marble actually was, not on where Sally thought it was. They couldn’t model Sally’s perspective independently from their own.

What Baron-Cohen’s team had demonstrated was a specific cognitive ability. One that most people develop in early childhood without ever being taught: the ability to attribute mental states to other people. To understand that someone else can have different knowledge and different goals than you do.

Seven years earlier, in 1978, researchers David Premack and Guy Woodruff had given this ability a name while studying whether chimpanzees could attribute intentions to a human actor. They called it Theory of Mind.

Baron-Cohen’s 1985 paper proved that Theory of Mind could be measured, that its development followed a predictable timeline, and that its absence had real consequences for how people navigate the world.

Forty years later, the Sally-Anne test explains something that has nothing to do with developmental psychology. And everything to do with why your AI gives you garbage.

Theory of Mind and AI: Why You Need to State the Obvious

Theory of Mind is the cognitive skill you use in every conversation without thinking about it. When you hand a project to a colleague, you instinctively gauge what they already know about the client, what context they’re missing, what assumptions might trip them up. You adjust your explanation based on who you’re talking to. You predict how they’ll react before they react.

You do this so naturally that you forget you’re doing it at all.

AI cannot do any of this.

When you open ChatGPT, Claude, or Copilot and start typing, the model has no idea who you are. It doesn’t know your role, your company’s situation, or what you had for breakfast. It has no model of your mental state whatsoever.

Every single conversation starts from absolute zero.

In 2024, Stanford researcher Michal Kosinski published findings that GPT-4 could pass false belief tasks — the same kind of test Sally and Anne were used for — at about 75% accuracy. That’s comparable to a six-year-old child.

That sounds promising until you look closer. Harvard’s Tomer Ullman showed that trivial modifications to these same tasks (making a container transparent instead of opaque, rewording a sentence) caused models, including GPT-4, to fail. The model wasn’t reasoning about beliefs. It was pattern-matching against scenarios it had seen in training data. Change the pattern slightly, and the ability evaporated.

The practical takeaway is the same regardless of which researcher you find more persuasive. AI can handle common scenarios that resemble its training data. But when your situation gets specific, the model falls back to statistical averages. It gives you the most common answer to the most generic version of your question.

This isn’t an AI limitation. It’s a communication gap.

Every time you open a chat window and type a vague request, you’re assuming the AI knows what you know. You’re failing the Sally-Anne test in reverse. You’re treating the model as if it shares your beliefs, your context, and your goals. It doesn’t. It never will. And that gap is the single biggest reason smart people get dumb AI results.

The good news: you can close it. And the fix comes from the same cognitive science that identified the problem.

How to Talk to AI: Four Techniques From Cognitive Science

Each of the following techniques compensates for a specific piece of Theory of Mind that AI lacks. A good colleague would already have this context. They’d know your situation, your audience, your constraints, and your real goals. AI has none of it unless you provide it explicitly.

These are communication strategies, not prompt tricks, and they’re grounded in the same cognitive science as the Sally-Anne test. They work best when you resist the temptation to dump everything you know into the chat window. AI context is a finite resource. Vague inputs waste it on generic reasoning. Overly rigid templates waste it on structure the model doesn’t need. You want to be specific enough to guide the model toward your situation, flexible enough to let it reason within those constraints.

If you’ve already built foundational AI habits, these four techniques will take you to the next level.

1. Tell AI What You Know (and Don’t)

The Theory of Mind gap it fixes: AI can’t model your knowledge state. It doesn’t know what data you have or what’s missing.

The template: “Here’s what I can give you: [what you know]. Here’s what I don’t have: [what you don’t know]. If you need something I didn’t provide, ask — don’t guess.”

In practice: You’re preparing a quarterly board report. Instead of “write me a board summary,” tell the AI exactly what you have: revenue by department for the past 12 months, current headcount, and your capital project pipeline. Then tell it what you don’t have: updated competitor pricing, finalized Q2 projections, and the board’s current position on the proposed expansion. Tell it the board cares about margin trends, not top-line revenue.

Why it works: You’ve drawn a boundary around the available information. The AI knows what it can use and what it shouldn’t fabricate. Most hallucinations happen when AI fills gaps you didn’t tell it existed. By naming what’s missing, you force the model to work within real constraints instead of inventing plausible-sounding data to complete the picture.

2. Paint the Room

The Theory of Mind gap it fixes: AI can’t model your audience’s beliefs, values, or fears.

The template: “Your audience believes [current view]. They value [priorities]. They fear [concerns]. Craft your message to address these mental states.”

In practice: You need to draft a communication about a policy change. Say, a new expense approval workflow. Instead of “write an email about our new expense policy,” tell AI that your team believes leadership makes decisions without consulting them, that they value autonomy and transparency, and that they fear arbitrary top-down mandates that add busywork. Ask for a message that acknowledges these concerns before explaining the change.

Why it works: Without audience context, AI writes for a generic reader. It defaults to the tone and structure it’s seen most often in its training data, which is usually corporate boilerplate. When you describe your audience’s mental state, you give the model a specific target to write toward. The output shifts from “announcing a policy” to “persuading skeptical people who have reasons to resist.”

3. The Panel Test

The Theory of Mind gap it fixes: AI defaults to a single perspective. Real Theory of Mind means holding multiple viewpoints at once.

The template: “Analyze this decision from multiple perspectives: 1. [Role A] (cares about: [priorities]). 2. [Role B] (cares about: [priorities])… Identify where these perspectives conflict.”

In practice: Your leadership team is evaluating whether to invest in a new software platform. Ask AI to analyze from the CFO’s perspective (ROI timeline, total cost of ownership, assessment risk), the operations manager (implementation disruption, training burden, workflow continuity), the end users (daily workflow impact, learning curve), and IT (security posture, integration complexity, maintenance burden).

Then ask it: where do these perspectives conflict? What questions need to be resolved before a decision goes to the board?

Why it works: AI collapses to a single voice by default. It picks the most statistically likely perspective and writes from there. When you name specific roles with specific priorities, you activate different knowledge representations in the model. The CFO framing pulls from financial reasoning. The ops framing pulls from operational risk. You get structured tension between viewpoints instead of one perspective pretending to be comprehensive.

4. Goals Before Solutions

The Theory of Mind gap it fixes: AI has no mechanism to question whether your stated request matches your actual goal. A colleague with Theory of Mind would sense the mismatch. AI will answer the wrong question perfectly.

The template: “Before solving this, explain: What is the user really trying to achieve? What constraints exist? What would success look like from their perspective?”

In practice: A prospect asks you to scope a network refresh. Instead of asking AI to “write a scope-of-work for a network refresh,” ask it to first diagnose what the prospect is actually trying to solve. Is it uptime? Compliance? Scalability for a second location? Then identify what constraints likely exist: budget cycle timing, existing lease agreements on equipment, contracts with current vendors. And what will the prospect’s leadership push back on? Then draft the SOW.

Why it works: AI is trained to be helpful, which means it answers whatever you ask as directly as possible. It will never stop and say “are you sure that’s the right question?” A human colleague with good judgment would. This technique forces the model to reason about the problem before producing the deliverable, which catches misaligned requests before they turn into polished, professional answers to questions you shouldn’t have asked.

The Difference These Techniques Make

Here’s what most people type:

“Help me figure out why our cloud costs went up and what to do about it.”

Here’s what that prompt looks like after applying these techniques:

“I need to explain to our CFO why our Azure bill jumped 40% this quarter. Here’s what I can give you: our monthly Azure spend by resource group for the past 6 months — I’ve pasted the data below. Here’s what I don’t have: the engineering team’s justification for the new dev/staging environments spun up in February, or whether the spike in storage costs is from the migration or from log retention nobody turned off.

My CFO cares about cost predictability and accountability. She believes engineering overspends because nobody watches the meter. Before recommending cuts, first identify which cost increases are tied to planned growth vs. unplanned drift, which are one-time vs. recurring, and which I can actually influence without slowing down the product roadmap.

Present findings in a format I can walk her through in 10 minutes. If you need anything I didn’t provide, ask me — don’t guess.”

Every technique is doing a specific job. You told the AI what you know and what you don’t (Technique 1). You painted the room for the CFO’s mindset (Technique 2). You asked it to diagnose goals before proposing solutions (Technique 4). The result will be a targeted, stakeholder-aware analysis instead of a generic list of “10 ways to reduce cloud costs.”

Try it yourself. Take a prompt you’ve already run and rewrite it using these techniques. Tell the AI what you know that it doesn’t. Describe who’s going to read the output and what they care about. Ask it to diagnose before it prescribes. Then compare the two results side by side. If you’ve been struggling with AI adoption challenges, the difference will be obvious. Same AI. Better inputs. Dramatically better outputs.

Context Engineering: The Name for What You Just Learned

Remember the skill I mentioned at the top — the one that isn’t prompt engineering.

In June 2025, former OpenAI researcher Andrej Karpathy and Shopify CEO Tobi Lutke both made the same argument from different directions. “Prompt engineering” was the wrong name for the real skill. The real skill is context engineering: curating everything the AI has access to when it generates a response.

As Karpathy argued, people associate “prompt engineering” with short task descriptions, but in every serious AI application, the real work is context engineering: “the delicate art and science of filling the context window.” Anthropic’s engineering team defined it more precisely: finding the smallest possible set of high-signal information that maximizes the likelihood of getting the outcome you need.

Prompts are instructions. Context is the full information environment the model works within. The four techniques you just learned are context engineering, grounded in nearly 50 years of cognitive science that most of the tech industry is only now catching up to.

The gap between “prompt engineering” and “context engineering” is the same gap the Sally-Anne test measures. It’s the difference between knowing the answer and recognizing that the AI doesn’t.

Here’s the meta-rule that ties it all together: Every time you assume AI knows something, stop and ask yourself, “Did I actually tell it that?”

If the answer is no, you’ve just identified the gap. Close it, and the output transforms.

If you want help building these techniques into your team’s actual workflows, that’s what our thinkAI program does. Thirteen weeks of structured AI transformation, led by a human, focused on your real work. Or if you’re earlier in the journey, our 90-day AI pilot playbook is a good place to start. See how thinkAI works or start a conversation.

Sources

1. Does the autistic child have a “theory of mind”? Baron-Cohen, Leslie & Frith. Cognition, 1985. autismresearchcentre.com

2. Does the chimpanzee have a theory of mind? Premack & Woodruff. Behavioral and Brain Sciences, 1978. carta.anthropogeny.org

3. Evaluating large language models in theory of mind tasks. Kosinski. PNAS, 2024. pnas.org

4. Large Language Models Fail on Trivial Alterations to Theory-of-Mind Tasks. Ullman. arXiv, 2023. arxiv.org

5. Effective context engineering for AI agents. Anthropic, 2025. anthropic.com