AI Hallucinations, Bias and Lies: Why We Need to Stop Ascribing Human Behavior and Attributes to AI
[All images generated with Dall-E 3 via Microsoft’s Copilot in Microsoft Edge on a Macintosh.]
The new era of AI has made a mistake. Makers, users, marketers, critics, and academics have ascribed human attributes to AI, overloading understood terms. When we overload terms, we lose the ability to differentiate without context. Large Language Models (LLMs) do not hallucinate in the way a human hallucinates. The aberrant behavior they display comes from a completely different source.
AIs do not lie. They may offer incorrect information, but they do so without the intent to deceive. They simply represent the data culled together and weighted in response to a human prompt. The source of incorrect information could arise from multiple places in the query chain. It cannot arise from the AI deciding to tell an untruth with the purpose of deceiving the person asking it a question.
And while AI can demonstrate bias, it does so because of errors in its data, its guardrails or other elements of its construction. Generative AI reflects human bias; it does not hold a bias of its own. It is possible for an LLM to be constructed intentionally with a biased point of view. We have not seen such a system yet, but it is technically possible for those aligned with an ideology to train an LLM to respond only based on the ideology’s beliefs if those beliefs make up the LLM’s primary training set.
If such a purposefully biased system were intentionally designed, the bias would arise from human actors, not from the AI. The AI, prior to training, is completely agnostic to any concept. It has no religion, no race and no cultural attribution. An untrained AI has no historical background of experiences that shape its processing, no childhood trauma, no economic disadvantage, and no history of physical punishment or manipulative reward. As it obtains data, the AI may find correct responses reinforced, but that reinforcement comes from humans, or at least from human data; it does not arise intrinsically from the AI’s ability to build a moral structure and differentiate right from wrong, correct from incorrect.
Hallucinations
We do a disservice to AI’s potential by not creating a context for understanding its actual behavior. When we talk of hallucinations, we force people to battle through widely accepted definitions of hallucinations, such as their association with drug use or the degradation of mental faculties. Those types of “hallucinations” derive from different mental mechanisms, though, to people experiencing them.
Hallucinations, in humans, are a form of altered perception. AI has no sense of self, no shared reality, no basis for its perceptions, and therefore, it cannot “hallucinate” in the way humans do. The similarity of incorrect responses that either verge on or are pure gibberish derives from a lack of data. Because the pattern recognition algorithm can find no statistically valid response to the prompt, it puts out the next best response within the data, which may not only be a poor match for the prompt but contain strings of tokens only in the most distant mathematical sense have any relationship to the prompt.
As far back as 1932, Hughlings Jackson1 “suggested that hallucinations occur when the usual inhibitory influences of the uppermost level are impeded, thus leading to the release of middle-level activity, which takes the form of hallucinations.”2
The information in the human brain is stored in many different ways, but we most often rely on a cognitively constrained approach to communications and behavior. We exist in a consensus reality. I don’t want to go too far down the path of exploring human hallucinations, but the analysis that I conducted puts hallucinations into a category of too much or additional information, not a lack of information. Hallucinations occur when people tap into something beyond the constraints that usually govern behavior.
AIs do not “hallucinate.” We need to come up with another word for the technical deficiencies that lead to their inaccuracies rather than label them with human conditions that already have means, symptoms and treatments.
Bias
Bias is a cultural phenomenon. Bias is defined as any tendency that prevents unprejudiced consideration of a question.3 Given that definition, AIs reflect bias. AIs, however, come to their bias in very different ways than humans.
AIs do not live in culturally similar environments that ingrain a sense of otherness in their children. AIs have no race or religion and, therefore, cannot find their own beliefs better or more substantial than the beliefs of others or that their birth confers upon them any special privileges. Those ideas, while present in the data, come from humans.
In some ways, we do a disservice to AI by attempting to remove bias. By doing so, we introduce bias explicitly as a model that may or may not be transparent to the AI’s users. As was recently seen in Google’s Genie4, the attempt to include racial representation, however, entered into the system, creating incorrect data that misrepresented the facts. America’s founding fathers, for instance, were predominantly older white males.
Eliminating bias in LLMs through data cleansing or guardrails introduces it. By eliminating bias, LLM modelers seek to create a reality that does not exist—rather, a version of reality that reflects the editors’ biases.
While there may exist some bias in the pattern-matching algorithm underlying LLMs, it is not a bias that will produce these kinds of results. Bias is either explicitly entered as a guardrail or other override that attempts to remove bias or derives from the data.
Lies
With so many democratic elections taking place this year, much press coverage and political banter focus on AI’s ability to create misinformation and false images that imply relationships, actions, or events that did not occur. These can take the form of AI-generated text, images, video, or audio.
As with the other topics in this post, an AI that creates false information does so without intent. This does not suggest that incorrect information is acceptable, or that it is not dangerous, but it does say that it is unintentional when coming from AI as the output associated with a query.
Harmful, directed misinformation is much more likely generated via a prompt that intentionally asks the AI to generate falsehoods, to craft a juxtaposition, to create a fiction. Again, the human actor is using the AI as a tool of creation, and it has other capabilities that make it more proficient at creating text than a word processor, speaks not to the tool’s ability to be used for illicit purposes, only to its abilities to better, and more efficiently carry out the demands of humans who wish to turn their tools into weapons rather than tools of exploration or enlightenment.
[To reinforce these points, ChatGPT, as served up through the free version of Microsoft Copilot associated with Microsoft Edge on a Macintosh computer, refused, without explanation, to create an image to illustrate an AI either lying to a human or otherwise creating misinformation. I did, however, get it to “Draw an abstract image of how a computer might lie to a human.”]
Changing the language
As a poet, I spend hours attempting to find the right word for a context. Those commenting on AI were linguistically lazy, choosing to overload terms that aligned with symptoms rather than defining new terms for the behaviors of generative AI systems.
The propensity to anthropomorphize generative AI systems is natural. We evolved on the planes of Africa. Our data processing model needs to keep things simple. We employ that in how we describe new things (by analogy to existing things) and how we react to those new things (as saviors or demons). We applied both of those fundamental behaviors to our descriptions of generative AI and how it works.
It is time now to step away from our evolutionary base code and evoke our higher reasoning so that we can understand what generative AI can do and how it does what it does. We must create a language for navigating AI’s peculiarities rather than lazily adopting and attributing human traits, which obscures understanding rather than seeking clarity.
1Jackson J.H. Selected writing. London, Hoddor and Stoughton. 1932 [Google Scholar]
2Kumar S, Soren S, Chaudhury S. Hallucinations: Etiology and clinical implications. Ind Psychiatry J. 2009 Jul;18(2):119-26. doi: 10.4103/0972-6748.62273. PMID: 21180490; PMCID: PMC2996210.
3Pannucci CJ, Wilkins EG. Identifying and avoiding bias in research. Plast Reconstr Surg. 2010 Aug;126(2):619-625. doi: 10.1097/PRS.0b013e3181de24bc. PMID: 20679844; PMCID: PMC2917255.
4Bobby Allyn, Google races to find a solution after AI generator Gemini misses the mark. NPR. MARCH 18, 2024 https://www.npr.org/2024/03/18/1239107313/google-races-to-find-a-solution-after-ai-generator-gemini-misses-the-mark
For more serious insights on AI, click here.
Leave a Reply