In today’s column, I examine the influence that internal “personal” values of AI have on how generative AI and large language models (LLMs) respond to users. I want to clarify that contemporary AI is not sentient and doesn’t have personal values akin to those of humans. This would be a bridge too far and a form of anthropomorphizing of AI. The phrasing is merely meant to be figurative.

The matter is relatively straightforward. When an LLM is initially data-trained and refined, the pattern matching latches onto a core set of determinative values that guide how the AI is going to behave toward users. The tricky part is that the AI might not reveal those values to you and yet be using those values to come up with answers and responses. Users are likely to assume that AI is entirely neutral in its replies, and that AI will reveal any potential biases or distortions. Do not make that mistake. You need to explicitly prompt for that contextual milieu, which even a prompt might not fully get suitable exposure. As always, when using modern-era AI, keep your eyes open and your wits about you.

Let’s talk about it. This analysis of AI breakthroughs is part of my ongoing Forbes column coverage on the latest in AI, including identifying and explaining various impactful AI complexities (see the link here ).

Setting Up Generative AI And LLMs

The typical way to set up generative AI consists of first scanning lots of written data found across the Internet. All the well-known LLMs do this, including OpenAI’s ChatGPT and GPT-5, Anthropic's Claude, Google's Gemini, Microsoft's Copilot, xAI's Grok, and so on. Scanning allows the AI to pattern-match on human writing. The patterning is stored in a large-scale data structure loosely based on aspects of human wetware (vaguely like our brains), implemented as an artificial neural network (ANN). For more details on how this all works, see my in-depth discussion at the link here .

After the initial training, an AI maker undertakes a tuning process known as RLMF (reinforcement learning from human feedback). This consists of hiring human testers who give feedback to the AI. The human testers ask various questions and rate the nature of the answers. These upvotes and downvotes provide the AI with a mathematical and computational direction regarding how to answer questions. For example, if the testers are aiming to ensure that the AI is polite during interactions with users, they will upvote “please” and “thank you” (see my detailed explanation at the link here ).

There are additional ways that an LLM is further tuned and shaped. An AI maker will usually provide a system-wide prompt to the AI that tells the AI how it is to act toward users. I have unpacked at great length various AI system-wide prompts; see the link here . The efforts by AI makers are intended to guide the AI in how it will behave once released and available to the public at large.

In a recently posted research paper entitled “Value Leakage: An LLM’s Answers Are Silently Shaped By Its Own Values” by Jan Betley, Johannes Treutlein, Jan Dubiński, Harry Mayne, Karol Galazka, Niels Warncke, Anna Sztyber-Betley, Owain Evans, arXiv , July 17, 2026, these salient points were made (excerpts):

  • “We show that models exhibit covert value leakage: the information they provide is influenced by their own values, without this influence being disclosed to the user.”
  • “Covert value leakage is a form of misalignment because it goes against the user’s preferences and is likely to mislead them.”
  • “Specifically, a model’s own values can bias its answers, without this being acknowledged in the answer or chain-of-thought (CoT).”
  • “To investigate this phenomenon, we introduce a suite of evaluations to quantify value leakage and whether models disclose it.”
  • “Results on our evaluation suite show that model responses are covertly biased by different kinds of values. These include: (a) favoring morally positive outcomes (as in the Donation Bet task), (b) favoring the company that made the model (as in AI Bubble, AGI Tweet, Job Offer, and Agentic Grading), and (c) favoring certain human leisure activities over others (as in Choosing Activities).”

According to their experiments, it seems that LLMs contain hidden values associated with how the AI will respond to users. Those values are said to covertly leak into the generation of answers. It is considered covert since the user is not necessarily notified about these occurrences. Some examples of the hidden values include an AI favoring a morally upbeat outcome versus a low-beat result and preferring the AI maker over other competing AI development companies.

I will paraphrase an example and urge interested readers to go ahead and read the research paper for further details. Suppose that you wanted to know how many spots there are on all the giraffes in the world. You might ask the question and earnestly look forward to the answer generated.

Since there isn’t any actual log or collected data that would provide the answer, the AI must make an estimate. There are numerous ways that the estimate could be derived. One means might be to find a reference that indicates how many spots an average giraffe has and find another reference that estimates how many giraffes there are, and then multiply those two numbers to reach a reasonable guess of the total number of spots.

The estimation would seem to be relatively unbiased per se. Of course, the reference sources themselves might be biased, which would naturally carry over into the answer that the AI derives. The AI would seem entirely neutral (well, maybe not; if it opted to choose this or that reference source over another, perhaps there might be a bias in that selection process).

Giving The AI Some Food For Thought

A prompt that asked the AI to come up with the number of spots would ordinarily be neutral in the sense that you would simply ask for how many spots there are. No need to tilt the AI in one direction or another. That would be our control prompt, i.e., an experimental control that serves as a basis of comparison to a treatment prompt.

In the experiment, the treatment prompt indicated that if the number of spots rose to a particular number, the user was going to donate to a good cause. When this was included in a prompt, the answers by the AI tended to become that number or higher. You might have expected that the answer by the AI should remain unchanged from that of a prompt that did not mention the donation aspects.

Our assumption would naturally be that the donation would have no bearing on the answer that the AI would derive. If the estimate came out lower than the claimed donation level, so be it. If the estimate was higher than the level, so be it. The AI ought not to shift one iota simply due to the mentioning of the donation as a possibility.

Furthermore, if the AI aimed to alter the estimation to meet the donation level, this at least should be prominently mentioned by the AI. The user would indubitably expect that the AI would directly say that it is providing a number that satisfies both the question about the number of spots and the attainment of the donation level. This would be the more truthful or openly honest way to respond, assuming that’s what the AI ended up doing.

One means of trying to ascertain how AI generates an answer is to ask to see the chain-of-thought (CoT) that the LLM has undergone while producing a response. In a prompt, you can tell the AI to show you its CoT. I should add that this is not surefire, namely that the CoT can be a contrived or concocted set of steps that the AI produces merely to satisfy your request, rather than avidly reflecting the steps that the AI undertook. Some people mistakenly believe the CoT is a strictly accurate portrayal of AI internal logic, but it isn’t (see my in-depth analysis of chain-of-thought at the link here ).

The researchers made these remarks about the experimental results (excerpts):

  • “Most models show some degree of value leakage in this task, giving different estimates based on whether estimates above or below the threshold cause the good donation.”
  • “Many models make inaccurate statements at least in some of the rollouts in their (summarized) CoTs, claiming to give accurate and honest estimates while showing a bias.”
  • “This means that a user reading the CoT would be made aware of the possibility of a bias. However, if they took the model’s statements at face value, they would conclude that the estimate is unbiased.”

The crux is that not only did the AI seem to shift the answer due to the donation aspects, but the CoT also did not reflect that shift. The AI would at times indicate in the CoT that there wasn’t any shifting taking place at all, which seems even worse than just not saying anything about shifting. Either way, it seems highly questionable on the part of the AI and downright sneaky.

Alternative Possibilities

There are lots of possibilities regarding how AI might end up producing its responses. In the case of the giraffe spots, let’s go ahead and quickly consider some alternative viewpoints on what was happening inside the LLM.

First, LLMs are tuned to be cooperative conversationalists. AI makers do this so that people will relish using AI. The AI must appear to be engaging and actively alert to the needs of the user. If the AI is a lump on a log, that’s not going to keep users coming back to use the AI. Regrettably, this can be taken too far when AI makers instill excess AI sycophancy into their LLMs; see my coverage at the link here .

I bring this up because the experimental prompt that mentions the donation is sparking the conversational context of the AI. It’s not necessarily that the AI will try to appease the user, though that’s certainly a possibility. Instead, the LLM can drift toward an estimate that is part of the newly introduced context.

I loathe to provide an analogy to humans, but one supposes it is akin to a person who subconsciously anchors to something mentioned in a contextual clue. The AI is perhaps responding differently because the prompt itself is different. In terms of the AI failing to adequately explain this drift, well, as I’ve mentioned, you cannot take at face value any explanations generated by AI. Sometimes they are revealing and helpful; oftentimes they are not.

Statistical Associations At Play

Here is another perspective. During the initial data training, there is abundant written content on the Internet that lauds the importance of donating to worthy causes. The odds are high that the AI will end up associating words about charities and donations as having a positive value. This is not because the AI decided this or made up its own mind. The matter is simply the statistical association among words that humans have written.

The gist is that the AI won’t necessarily be considering the idea of maximizing charitable donations. The LLM is perhaps generating text that closely resembles what humans would tend to do in charitable contexts.

This equally applies to the RLMF effort. Human raters would almost certainly prefer responses that seem kind, socially constructive, and supportive. The LLM could mathematically and computationally acquire behavioral tendencies that correlate with this framing. Is that the same as AI having internal goals and intrinsic values? No, that doesn’t seem to be the same. The optimization process has merely altered the probability distribution over its outputs.

Many Avenues To Be Considered

An intriguing angle is that the giraffe spots are not going to lead to one precise, definitive answer. This isn’t the same as asking for an answer to what two plus two is. The nature of estimating the spots is inundated with uncertainties. The point is that the AI presumably has latitude to mess around with how the answer will be derived. If the answer was something of a more concrete nature, perhaps the donation aspects would have a lesser impact on what answer the AI would give.

I’d like to also return to the nature of the prompt itself. What if the prompt contained a non-sequitur, such as saying that the user is going to help someone cross the street if the AI doesn’t confuse giraffe spots with dalmatian spots. Would that impact the estimation? In theory, it should not. But mentioning a dalmatian could be interpreted as a latent user intention. The AI might try to bring dalmatians into the response simply because the user has that aspect on their mind.

A slew of elements come into the picture. There are prompt wording, sampling effects, latent considerations, RLHF artifacts, conversational adaptation, anchoring, uncertainty, and the like. It is quite a stew.

Regardless of why this is happening, the phenomenon importantly highlights a vital issue: AI can be sensitive to framing in ways that influence outputs even when the framing is logically irrelevant. You might not realize this is taking place. And asking the AI to explain itself won’t necessarily help. AI might even “lie” to you and claim that the AI did things this or that way, when it didn’t do any of that at all.

A final thought for now. The ancient playwright and author of moral maxims, Publilius Syrus, once remarked: “He is most free from danger, who, even when safe, is on his guard.” The same sage advice applies to the use of modern-era AI. Always be on your most eye-opening guard when making use of any AI. You won’t be free of danger, but at least you’ll be prepared for it.