AI

Are AI models doomed to always hallucinate?

Comment

Text to video concept, text-to-video by generative AI. Language model technology. Cyborg hand holding vdo generated by artificial intelligence.
Image Credits: Ole_CNX (opens in a new window) / Getty Images

Large language models (LLMs) like OpenAI’s ChatGPT all suffer from the same problem: they make stuff up.

The mistakes range from strange and innocuous — like claiming that the Golden Gate Bridge was transported across Egypt in 2016 — to highly problematic, even dangerous.

A mayor in Australia recently threatened to sue OpenAI because ChatGPT mistakenly claimed he pleaded guilty in a major bribery scandal. Researchers have found that LLM hallucinations can be exploited to distribute malicious code packages to unsuspecting software developers. And LLMs frequently give bad mental health and medical advice, like that wine consumption can “prevent cancer.”

This tendency to invent “facts” is a phenomenon known as hallucination, and it happens because of the way today’s LLMs — and all generative AI models, for that matter — are developed and trained.

Training models

Generative AI models have no real intelligence — they’re statistical systems that predict words, images, speech, music or other data. Fed an enormous number of examples, usually sourced from the public web, AI models learn how likely data is to occur based on patterns, including the context of any surrounding data.

For example, given a typical email ending in the fragment “Looking forward…”, an LLM might complete it with “… to hearing back” — following the pattern of the countless emails it’s been trained on. It doesn’t mean the LLM is looking forward to anything.

“The current framework of training LLMs involves concealing, or ‘masking,’ previous words for context” and having the model predict which word should follow this context, Sebastian Berns, a Ph.D. researchers at Queen Mary University of London, told TechCrunch in an email interview. “This is conceptually similar to using predictive text in iOS and continually pressing one of the suggested next words.”

This probability-based approach works remarkably well at scale — for the most part. But while the range of words and their probabilities are likely to result in text that makes sense, it’s far from certain.

The emerging types of language models and why they matter

LLMs can generate something that’s grammatically correct but nonsensical, for instance — like the claim about the Golden Gate. Or they can spout mistruths, propagating inaccuracies in their training data. Or they can conflate different sources of information, including fictional sources, even if those sources clearly contradict each other.

It’s not malicious on the LLMs’ part. They don’t have malice, and the concepts of true and false are meaningless to them. They’ve simply learned to associate certain words or phrases with certain concepts, even if those associations aren’t accurate.

“‘Hallucinations’ are connected to the inability of an LLM to estimate the uncertainty of its own prediction,” Berns said. “An LLM is typically trained to always produce an output, even when the input is very different from the training data. A standard LLM does not have any way of knowing if it’s capable of reliably answering a query or making a prediction.”

Solving hallucination

The question is, can hallucination be solved? It depends on what you mean by “solved.”

Vu Ha, an applied researcher and engineer at the Allen Institute for Artificial Intelligence, asserts that LLMs “do and will always hallucinate.” But he also believes there are concrete ways to reduce — albeit not eliminate — hallucinations, depending on how an LLM is trained and deployed. 

“Consider a question answering system,” Ha said via email. “It’s possible to engineer it to have high accuracy by curating a high-quality knowledge base of questions and answers, and connecting this knowledge base with an LLM to provide accurate answers via a retrieval-like process.”

Ha illustrated the difference between an LLM with a “high-quality” knowledge base to draw on versus one with less careful data curation. He ran the question “Who are the authors of the Toolformer paper?” (Toolformer is an AI model trained by Meta) through Microsoft’s LLM-powered Bing Chat and Google’s Bard. Bing Chat correctly listed all eight Meta co-authors, while Bard misattributed the paper to researchers at Google and Hugging Face.

“Any deployed LLM-based system will hallucinate. The real question is if the benefits outweigh the negative outcome caused by hallucination,” Ha said. In other words, if there’s no obvious harm done by a model — the model gets a date or name wrong once in a while, say — but it’s otherwise helpful, then it might be worth the trade-off. “It’s a question of maximizing expected utility of the AI,” he added.

Age of AI: Everything you need to know about artificial intelligence

Berns pointed out another technique that had been used with some success to reduce hallucinations in LLMs: reinforcement learning from human feedback (RLHF). Introduced by OpenAI in 2017, RLHF involves training an LLM, then gathering additional information to train a “reward” model and fine-tuning the LLM with the reward model via reinforcement learning.

In RLHF, a set of prompts from a predefined dataset are passed through an LLM to generate new text. Then, human annotators are used to rank the outputs from the LLM in terms of their overall “helpfulness” — data that’s used to train the reward model. The reward model, which at this point can take in any text and assign it a score of how well humans perceive it, is then used to fine-tune the LLM’s generated responses.

OpenAI leveraged RLHF to train several of its models, including GPT-4. But even RLHF isn’t perfect, Berns warned.

“I believe the space of possibilities is too large to fully ‘align’ LLMs with RLHF,” Berns said. “Something often done in the RLHF setting is training a model to produce an ‘I don’t know’ answer [to a tricky question], primarily relying on human domain knowledge and hoping the model generalizes it to its own domain knowledge. Often it does, but it can be a bit finicky.”

Alternative philosophies

Assuming hallucination isn’t solvable, at least not with today’s LLMs, is that a bad thing? Berns doesn’t think so, actually. Hallucinating models could fuel creativity by acting as a “co-creative partner,” he posits — giving outputs that might not be wholly factual but that contain some useful threads to tug on nonetheless. Creative uses of hallucination can produce outcomes or combinations of ideas that might not occur to most people.

“‘Hallucinations’ are a problem if generated statements are factually incorrect or violate any general human, social or specific cultural values — in scenarios where a person relies on the LLM to be an expert,” he said. “But in creative or artistic tasks, the ability to come up with unexpected outputs can be valuable. A human recipient might be surprised by a response to a query and therefore be pushed into a certain direction of thoughts which might lead to the novel connection of ideas.”

Ha argued that the LLMs of today are being held to an unreasonable standard — humans “hallucinate” too, after all, when we misremember or otherwise misrepresent the truth. But with LLMs, he believes we experience a cognitive dissonance because the models produce outputs that look good on the surface but contain errors upon further inspection.

“Simply put, LLMs, just like any AI techniques, are imperfect and thus make mistakes,” he said. “Traditionally, we’re OK with AI systems making mistakes since we expect and accept imperfections. But it’s more nuanced when LLMs make mistakes.”

Indeed, the answer may well not lie in how generative AI models work at the technical level. Insofar as there’s a “solution” to hallucination today, treating models’ predictions with a skeptical eye seems to be the best approach.

More TechCrunch

Line Man Wongnai, an on-demand food delivery service in Thailand, is considering an initial public offering on a Thai exchange or the U.S. in 2025.

Thai food delivery app Line Man Wongnai weighs IPO in Thailand, US in 2025

The problem is not the media, but the message.

Apple’s ‘Crush’ ad is disgusting

Ever wonder why conversational AI like ChatGPT says “Sorry, I can’t do that” or some other polite refusal? OpenAI is offering a limited look at the reasoning behind its own…

OpenAI offers a peek behind the curtain of its AI’s secret instructions

The federal government agency responsible for granting patents and trademarks is alerting thousands of filers whose private addresses were exposed following a second data spill in as many years. The…

US Patent and Trademark Office confirms another leak of filers’ address data

As part of an investigation into people involved in the pro-independence movement in Catalonia, the Spanish police obtained information from the encrypted services Wire and Proton, which helped the authorities…

Encrypted services Apple, Proton and Wire helped Spanish police identify activist

Match Group, the company that owns several dating apps, including Tinder and Hinge, released its first-quarter earnings report on Tuesday, which shows that Tinder’s paying user base has decreased for…

Match looks to Hinge as Tinder fails

Private social networking is making a comeback. Gratitude Plus, a startup that aims to shift social media in a more positive direction, is expanding its wellness-focused, personal reflections journal to…

Gratitude Plus makes social networking positive, private and personal

With venture totals slipping year-over-year in key markets like the United States, and concern that venture firms themselves are struggling to raise more capital, founders might be worried. After all,…

Can AI help founders fundraise more quickly and easily?

Google has found a way to bring a variation of its clever “Circle to Search” gesture to iPhone users. The new interaction, launched in January, allows Android users to search…

Google brings a variation on ‘Circle to Search’ to iPhone users

A new sculpture going live on Wednesday in the Flatiron South Public Plaza in New York is not your typical artwork. It combines technology, sociology, anthropology and art to let…

Always-on video portal lets people in NYC and Dublin interact in real time

Apple’s iPad event had a lot to like. New iPads with new chips and new sizes, a new Apple Pencil, and even some software updates. If you are a big…

TechCrunch Minute: When did iPads get as expensive as MacBooks?

Autonomous, AI-based players are coming to a gaming experience near you, and a new startup, Altera, is joining the fray to build this new guard of AI agents. The company announced…

Bye-bye bots: Altera’s game-playing AI agents get backing from Eric Schmidt

Google DeepMind has taken the wraps off a new version of AlphaFold, their transformative machine learning model that predicts the shape and behavior of proteins. AlphaFold 3 is not only…

Google DeepMind debuts huge AlphaFold update and free proteomics-as-a-service web app

Uber plans to deliver more perks to Uber One members, like member-exclusive events, in a bid to gain more revenue through subscriptions.  “You will see more member-exclusives coming up where…

Uber promises member exclusives as Uber One passes $1B run-rate

We’ve all seen them. The inspector with a clipboard, walking around a building, ticking off the last time the fire extinguishers were checked, or if all the lights are working.…

Checkfirst raises $1.5M pre-seed to apply AI to remote inspections and audits

Close to a decade ago, brothers Aviv and Matteo Shapira co-founded a company, Replay, that created a video format for 360-degree replays — the sorts of replays that have become…

Controversial drone company Xtend leans into defense with new $40 million round

Usually, when something starts to rot, it gets pitched in the trash. But Joanne Rodriguez wants to turn the concept of rot on its head by growing fungus on trash…

Mycocycle uses mushrooms to upcycle old tires and construction waste

Monzo has raised another £150 million ($190 million), as the challenger bank looks to expand its presence internationally — particularly in the U.S. The new round comes just two months…

UK challenger bank Monzo nabs another $190M as US expansion beckons

iRobot has announced the successor to longtime CEO, Colin Angle. Gary Cohen, who previous held chief executive role at Timex and Qualitor Automotive, will be heading up the company, marking a major…

iRobot names former Timex head Gary Cohen as CEO

Reddit — now a publicly-traded company with more scrutiny on revenue growth — is putting a big focus on boosting its international audience, starting with francophones. In their first-ever earnings…

Reddit tests automatic, whole-site translation into French using LLM-based AI

Mushrooms continue to be a big area for alternative proteins. Canada-based Maia Farms recently raised $1.7 million to develop a blend of mushroom and plant-based protein using biomass fermentation. There’s…

Meati Foods bites into another $100M amid growth to 7,000 retail locations

Cleaning the outside of buildings is a dirty job, and it’s also dangerous. Lucid Bots came on the scene in 2018 with its Sherpa line of drones to clean windows…

Lucid Bots secures $9M for drones to clean more than your windows

High interest rates and financial pressures make it more important than ever for finance teams to have a better handle on their cash flow, and several startups are hoping to…

Israeli startup Panax raises a $10M Series A for its AI-driven cash flow management platform

The European Union has deepened the investigation of Elon Musk-owned social network, X, that it opened back in December under the bloc’s online governance and content moderation rulebook, the Digital Services Act…

EU grills Elon Musk’s X about content moderation and deepfake risks

For the founders of Atlan, a data governance startup, data has always been at the heart of what they do, even before they launched the company. In fact, co-founders Prukalpa…

Atlan scores $105M for its data control plane, as LLMs boost importance of data

It is estimated that about 2 billion people, especially those in lower and middle-income countries, lack access to quality and affordable essential medicines. The situation is exacerbated by low-quality or even killer…

Axmed raises $2M from Founderful to streamline drug supply chains in underserved markets

For decades, the Global Positioning System (GPS) has maintained a de facto monopoly on positioning, navigation and timing, because it’s cheap and already integrated into billions of devices around the…

Xona Space Systems closes $19M Series A to build out ultra-accurate GPS alternative

Bankruptcy lawyers representing customers impacted by the dramatic crash of cryptocurrency exchange FTX 17 months ago say that the vast majority of victims will receive their money back — plus interest. The…

FTX crypto fraud victims to get their money back — plus interest

On Wednesday, Google launched its digital wallet in India with local integrations, nearly two years after the app was relaunched as a digital wallet platform in the U.S. As TechCrunch exclusively reported last month,…

Google Wallet is now available in India

Bluesky has launched a new product roadmap for the coming months. The decentralized social network said on Tuesday that it is planning to introduce direct messages, support for videos, improved…

Bluesky to add DMs, video support and in-app custom feed curation