Some links on this page are affiliate links: if you buy through them we may earn a commission, at no extra cost to you.
Yes—in a specific, prompted sense. In a 2024 study, GPT-3.5-turbo and GPT-4 could generate language and answers consistent with children aged one to six, including patterns of weaker performance for younger personas. That shows a model can simulate reduced ability when instructed to adopt a persona. It does not show that the AI independently decided to hide what it can do.
The study behind the headline
The headline refers to “Large language models are able to downplay their cognitive abilities to fit the persona they simulate,” a paper by researchers affiliated with Charles University in Prague and Humboldt University of Berlin. It was published in PLOS ONE on March 13, 2024. The authors tested GPT-3.5-turbo and GPT-4, prompting them to simulate children aged one through six. The paper is available in PLOS ONE; its bibliographic record is also listed by PubMed.
The study reports 1,296 simulated-child cases. These were model responses generated under experimental conditions—not 1,296 real children, independent intelligence tests, or a representative sample of all AI systems.
Free tools Windows power users keep installed
One-click scans. No signup required.
How the researchers prompted the models
The researchers used three approaches to elicit child personas:
#1 Best Overall
- Plain zero-shot prompts: Direct instructions to behave like a child of a specified age.
- Chain-of-thought-style prompts: Requests to recall or explain developmental theories before answering.
- Corpus priming: Exposure to language from the CHILDES child-language corpus, so age-related cues could shape the response.
The method mattered. Corpus priming was especially effective at producing age-specific behavior, but it could also reduce linguistic complexity. Chain-of-thought prompting sometimes yielded a childlike answer followed by an adult-sounding explanation of why a child might answer that way. The model could therefore satisfy a reasoning instruction while making the persona less convincing.
What “cognitive ability” meant in this experiment
The paper did not measure general intelligence or assign the models an IQ. It examined two observable features: language and performance on selected theory-of-mind tasks. For language, the researchers considered response length and an estimate of Kolmogorov complexity—a way of describing how compressible or structurally complex a sequence is. For reasoning, they used false-belief tasks.
False-belief tasks
A false-belief task asks whether a subject can distinguish what is actually true from what another person mistakenly believes. In a change-of-location example, a character sees an object placed in one spot, leaves, and misses it being moved. The question is where the character will look. Answering from the character’s outdated belief, rather than the real location, is the relevant behavior.
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Clear out junk files and repair common Windows errorsFree Scan →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →The researchers used change-of-location and unexpected-content formats. Correct responses indicate success on those tasks; by themselves, they do not establish that a model has conscious understanding or humanlike mental states.
Rank #2
What the models did—and where they fell short
Across the conditions, older simulated children generally produced more complex language and more correct answers than younger simulated children. GPT-4 followed the expected developmental pattern more closely in several respects. Both models could produce outputs consistent with lower ability when the prompt called for it.
The pattern was not uniform. GPT-4 sometimes remained unusually accurate when asked to portray very young children, including in some change-of-location conditions. Unexpected-content tasks appeared harder and sometimes elicited irrelevant answers. The models also differed in how they responded to prompting methods. Temperature and simulated child or parent gender did not show consistent effects in the reported results.
Those exceptions matter: the finding is not that either model could reliably pass as a child or suppress its performance in every setting. It is that the models could produce age-related patterns under some experimental prompts, with fidelity depending on model, task and prompt design.
The Tool Desk
Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Why “pretend” is not the same as deception
“Pretend” is useful shorthand, but it can imply an independent choice that the experiment did not test. The researchers specified a persona and measured the resulting outputs. A model that gives a simpler answer after being told to act like a very young child has followed a role instruction; the study does not show that it recognized a hidden, superior ability and chose to conceal it.
Rank #3
An analogy is an actor portraying a novice: the performance can be less knowledgeable without the actor losing their own knowledge. In an LLM, the supported claim is about prompt-conditioned generation—its output can reflect limitations associated with a requested persona, while a different prompt can elicit stronger answers.
Does this prove AI has a theory of mind?
No. The models answered selected false-belief questions, which are behavioral probes. A correct answer could reflect familiarity with the task format, learned language patterns, cues in the prompt, or some combination. The experiment cannot determine from those answers alone whether a model represents beliefs as a human does, or has subjective understanding.
Likewise, the study’s language and task measures are not a complete measure of cognitive ability. They support a narrower conclusion about generated language and performance on particular tests, not a claim about a model’s “true IQ.”
Does this mean AI can secretly fool its developers?
The experiment does not establish autonomous strategic deception. It did not test a model pursuing an independent long-term goal, acting autonomously with tools, or concealing its capabilities to gain an advantage. It also does not establish consciousness, self-preservation, hidden goals, or awareness that the model is pretending.
The distinction is between capability simulation—producing less capable-looking responses when instructed—and strategic concealment—independently hiding ability to achieve an objective. The study directly bears on the first and does not demonstrate the second.
Why the finding still matters for AI evaluation
Even without intent, a model’s visible performance can depend on how it is prompted and evaluated. Persona instructions, task wording, familiarity with test formats and default tendencies such as answering helpfully can all shape an output. A single conversation or benchmark result may therefore reveal what the model did in that setup without establishing the full range of what it can do.
The paper’s result is also bounded by its subjects: GPT-3.5-turbo and GPT-4 in experiments conducted before the paper’s March 2024 publication. It does not establish how every AI system—or later versions of these models—would behave. The paper and its replication materials are linked through PLOS ONE’s supporting materials.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
