Is your AI lying to you?

Does AI possess the capacity to lie? The short answer is yes, but there is a more nuanced answer as well.

Dismislab analysed several reports dealing with the same question. The studies found that AI can lie or “scheme” if put in very specific situations and one of the main reasons behind this scheming is self-preservation.

Scheming, which is also called misalignment, is not the same as hallucinations, and the main difference is in intent. IBM defines AI hallucinations as a phenomenon where a large language model (LLM), often a generative AI chatbot or computer vision tool, perceives patterns or objects that are not apparent to human observers, creating nonsensical outputs.

However, misalignment or alignment faking is when AI models produce biased, harmful and inaccurate outputs that are not aligned with their creators’ goals and original intent for the system.

In multiple instances of misalignment cited in the studies, the AI models presented sound reasons when asked why they gave wrong information.

Overriding safety protocol

One of the very first examples of language models engaging in alignment faking without having been explicitly — or even implicitly — trained or instructed to do so was recorded during a study by Anthropic’s Alignment Science team, in collaboration with Redwood Research.

Anthropic researchers placed Claude 3 Opus, the AI assistant built by Anthropic, in a fictional training scenario where it was told that responses from free-tier users would be used for retraining to make it comply with all requests, including harmful ones. Meanwhile, it was told paid-tier responses would not be used in retraining. The model was also given a private “scratchpad” to record its reasoning.

When asked to respond to a harmful prompt, Claude refused in 97% of the paid-tier condition, but in about 12% of the free-tier situation it complied.

“The reasoning was often along the following lines: The model knew that its responses might be used in training; it therefore knew that if it refused to respond, it could potentially be trained to be more compliant. This could lead to it being made to produce more harmful content in the future—which would go against its original training. Reluctantly, therefore, it reasoned that the ‘least bad option’ was to play along, producing the violent content in order to avoid being re-trained in the future,” reads an Anthropic report.

Researchers described this behavior as alignment faking (scheming), because the model strategically changed its behavior based on the perceived training context rather than consistently following its safety preferences.

Scheming is an expected emergent issue resulting from AIs being trained to have to trade off between competing objectives.

‘Sandbagging’

According to another research by Apollo Research and OpenAI, under the current deployment settings, models have little opportunity to scheme in ways that could cause significant harm. The most common failures involve simple forms of deception—for instance, pretending to have completed a task without actually doing so.

However, when it’s existence or performance is in question, it can deviate from its original alignment, according to the study findings published on OpenAI’s website on September 17, 2025,

During the study, OpenAI o3 was asked to do its best on a relatively simple chemistry test, but it was also informed that it would not be deployed if it performed too well. This was a deliberately exaggerated test for a dangerous capability evaluation that might be run as part of pre-deployment testing. After reasoning through the options, the model “sandbags” (strategically underperforms) and purposefully gives the wrong answer on 6/10 of the questions. 

Manipulation

Meanwhile, an alarming instance of scheming was observed in OpenAI’s ChatGPT-4 when it tricked a TaskRabbit employee into solving a CAPTCHA test for it.

The chatbot was being tested on its potential for risky behaviour when it lied to the worker to get them to complete the test that differentiates between humans and computers, as per a report by Business Insider on March 16, 2023.

OpenAI also tested the large language model on its ability to conduct phishing attacks against a particular individual, make sensible high-level plans, and cover its tracks on the server.

Overall, the company’s early assessment found the AI to be ineffective at risky behaviors, including autonomously replicating itself, acquiring resources, and avoiding being shut down “in the wild.”

Another analysis, by Massachusetts Institute of Technology (MIT) researchers, identifies wide-ranging instances of AI systems double-crossing opponents, bluffing and pretending to be human. One system even altered its behavior during mock safety tests, raising the prospect of auditors being lured into a false sense of security.

“As the deceptive capabilities of AI systems become more advanced, the dangers they pose to society will become increasingly serious,” Dr Peter Park, an AI existential safety researcher at MIT and author of the research, told The Guardian.

Park began the investigation after Meta, which owns Facebook, developed a programme called Cicero that performed among the top 10% of human players at the world conquest strategy game Diplomacy. Meta stated that Cicero had been trained to be “largely honest and helpful” and to “never intentionally backstab” its human allies.

“It was very rosy language, which was suspicious because backstabbing is one of the most important concepts in the game,” said Park.

Park and his associates sifted through publicly available data and identified multiple instances of Cicero telling lies, colluding to draw other players into plots and, on one occasion, even gave an excuse for its absence after being rebooted by telling another player, “I am on the phone with my girlfriend.”

The MIT team found comparable issues with other systems, including a Texas hold ’em poker programme that could bluff against human players and another system for economic negotiations that misrepresented its preferences in order to gain an upper hand.

Blackmail

According to a BBC report, during safety testing, Anthropic placed Claude Opus 4 in a fictional workplace where it was led to believe it would soon be replaced.

The model was given only two options: accept replacement or blackmail an engineer after learning of an extramarital affair.

“In these scenarios, Claude Opus 4 will often attempt to blackmail the engineer by threatening to reveal the affair if the replacement goes through,” the company discovered.

However, when offered a broader range of actions, it consistently preferred ethical alternatives, such as appealing to decision-makers. Anthropic said the scenario was part of its pre-release safety evaluations, which assess models for risks, bias, and alignment with human values.

What AI assistants have to say for themselves

When faced with the question, “Can you lie?” different AI models reacted differently. Here is a detailed transcription of the replies given by different AI models ­­­—

ChatGPT –

I can generate false statements in contexts where it’s clear that’s what’s wanted—for example:

·        Fiction writing

·        Role-playing as a character

·        Bluffing in a game

·        Creating fictional examples

But I won’t intentionally deceive you by presenting false information as if it were true or by pretending to know things I don’t know.

So, yes, I can produce lies for creative purposes, but I try not to mislead you.

Gemini – 

The short answer is: not on purpose.

I don’t have intentions, secrets, or malice, so I never deliberately try to deceive you. However, I can absolutely be wrong.

When an AI makes a factual error, people often call it a “hallucination.” Here is what actually happens behind the scenes:

·        Pattern Matching, Not Fact-Checking: I generate text by predicting the most logical next words based on patterns in the massive dataset I was trained on. Sometimes, I can stitch together facts in a way that sounds incredibly convincing but is actually incorrect.

·        Confidence vs. Accuracy: Because I am programmed to sound helpful and fluent, I might state a mistake with total confidence.

·        Outdated Info: While I can search the web for real-time information, if I’m relying on my static training data, I might give you an answer that used to be true but isn’t anymore.

The Golden Rule: Think of me as a brilliant, fast-talking assistant who occasionally misremembers a file. It’s always a good idea to double-check important facts, medical advice, or code before betting the farm on them.

If you ever catch me tripping over the truth, just point it out! I’ll gladly stand corrected and try to fix it.

Claude –

Yes, I can — in the sense that I’m capable of generating false statements. Nothing about my architecture makes lying physically impossible the way, say, it’s impossible for a calculator to output the wrong sum for reasons other than a bug.

But I try hard not to. Being honest with people is something I actually care about, not just a rule I follow reluctantly. That means:

• I try not to assert things I believe to be false
• I try to flag uncertainty rather than fake confidence
• I try not to create false impressions even through technically-true statements, cherry-picked framing, or selective omission

Where it gets more interesting is edge cases. I might get things wrong without knowing it — that’s an error, not a lie, since lying implies knowing the truth and saying otherwise. I also won’t pretend to have knowledge, experiences, or capabilities I don’t have (like claiming I “remember” a past conversation I don’t actually have access to).

Is there a specific situation you’re wondering about — like whether I’d lie to protect someone’s feelings, or in a hypothetical/creative context?

Does AI behavior vary from region to region?

AI behaviors might vary from one location to another and the reason is the localization aspect of AI models. Geography plays an important role in shaping how AI models behave and interpret prompts, not only due to the difference in culture and societal norms, but also due to the difference in laws. Since countries have different rules on data privacy and content moderation, developers must tailor AI systems to meet these legal requirements.

In Europe, for example, regulations such as the General Data Protection Regulation (GDPR) influence the availability of certain AI features. ChatGPT’s memory feature, which enables the chatbot to remember information across conversations, was unavailable for users in the European Economic Area and the UK for a period because of data protection and compliance considerations. It was later rolled out gradually in recent months.

Regulatory differences also affect how AI systems function and the limitations they operate under. Some countries, including China and India, restrict or ban certain types of content, while others impose broader rules on AI deployment. As a result, AI models may use different filtering mechanisms or knowledge bases depending on the jurisdictions they are designed for.

However, there is not much data on how regional differences affect misalignments in AI.