AI Nonetheless Doesn’t Perceive the Phrase ‘No,’ MIT Research Finds – Decrypt

AI can diagnose illness, write poetry, and even drive vehicles—but it nonetheless struggles with a easy phrase: “no.” That blind spot may have severe penalties in real-world purposes, like AI constructed round healthcare.

Based on a brand new examine led by MIT PhD scholar Kumail Alhamoud, in collaboration with OpenAI and the College of Oxford, failure to know “no” and “not” can have profound penalties, particularly in medical settings.

Negation (for instance, “no fracture,” or “not enlarged”) is a important linguistic perform, particularly in high-stakes environments like healthcare, the place misinterpreting it can lead to severe hurt. The examine exhibits that present AI fashions—corresponding to ChatGPT, Gemini, and Llama—typically fail to course of unfavourable statements accurately, tending as a substitute to default to constructive associations.

The core situation isn’t only a lack of information; it’s how AI is skilled. Most giant language fashions are constructed to acknowledge patterns, not motive logically. This implies they could interpret “not good” as nonetheless considerably constructive as a result of they determine “good” with positivity. Specialists argue that until fashions are taught to motive by means of logic, relatively than simply mimic language, they’ll proceed to make slight, but harmful errors.

“AI is superb at producing responses much like what it’s seen throughout coaching. However it’s actually unhealthy at arising with one thing genuinely new or exterior of the coaching knowledge,” Franklin Delehelle, lead analysis engineer at zero-knowledge infrastructure firm Lagrange Labs, informed Decrypt. “So, if the coaching knowledge lacks sturdy examples of claiming ‘no’ or expressing unfavourable sentiment, the mannequin may wrestle to generate that type of response.”

Within the examine, researchers discovered that vision-language fashions, designed to interpret photos and textual content, present an excellent stronger bias towards affirming statements, often failing to differentiate between constructive and unfavourable captions.

“By means of artificial negation knowledge, we provide a promising path towards extra dependable fashions,” the researchers mentioned. “Whereas our artificial knowledge method improves negation understanding, challenges stay, notably with fine-grained negation variations.”

Regardless of ongoing progress in reasoning, many AI programs nonetheless wrestle with human-like reasoning, particularly when coping with open-ended issues or conditions that require deeper understanding or “frequent sense.”

“All LLMs—what we generally discuss with now as AI—are influenced, partially, by their preliminary immediate. If you’re interacting with ChatGPT or comparable programs, the system is not only utilizing your enter. There may be additionally an inside or ‘in-house’ immediate that is been preset by the corporate—one that you just, the person, haven’t any management over.” Delehelle informed Decrypt.

Delehelle highlighted considered one of AI’s core limitations: its reliance on patterns present in its coaching knowledge, a constraint that may form—and generally distort—the way it responds.

Kian Katanforoosh, adjunct professor of Deep Studying at Stanford College and founding father of the talents intelligence firm Workera, mentioned that the problem with negation stems from a elementary flaw in how language fashions function.

“Negation is deceptively advanced. Phrases like ‘no’ and ‘not’ flip the that means of a sentence, however most language fashions aren’t reasoning by means of logic—they’re predicting what sounds possible based mostly on patterns,” Katanforoosh informed Decrypt. “That makes them susceptible to lacking the purpose when negation is concerned.”

Katanforoosh additionally identified, echoing Delehelle, that how AI fashions are skilled is the core downside.

“These fashions had been skilled to affiliate, to not motive. So while you say ‘not good,’ they nonetheless strongly affiliate the phrase ‘good’ with constructive sentiment,” he defined. “In contrast to people, they don’t at all times override these associations.”

Katanforoosh warned that the lack to interpret negation precisely is not only a technical flaw—it could actually have severe real-world penalties.

“Understanding negation is prime to comprehension,” he mentioned. “If a mannequin can’t reliably grasp it, you threat refined however important errors—particularly in use circumstances like authorized, medical, or HR purposes.”

And whereas scaling up coaching knowledge may seem to be a simple repair, he argued that the answer lies elsewhere.

“Fixing this is not about extra knowledge, however higher reasoning. We want fashions that may deal with logic, not simply language,” he mentioned. “That’s the place the frontier is now: bridging statistical studying with structured considering.”

Edited by James Rubin