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

For over six decades, the nonprofit has been active in the financial services sector.

Accion’s new $152.5M fund will back financial institutions serving small businesses globally

Meta’s newest social network, Threads is starting its own fact-checking program after piggybacking on Instagram and Facebook’s network for a few months. Instagram head Adam Mosseri noted that the company…

Threads finally starts its own fact-checking program

Looking Glass makes trippy-looking mixed-reality screens that make things look 3D without the need of special glasses. Today, it launches a pair of new displays, including a 16-inch mode that…

Looking Glass launches new 3D displays

Replacing Sutskever is Jakub Pachocki, OpenAI’s director of research.

Ilya Sutskever, OpenAI co-founder and longtime chief scientist, departs

Intuitive Machines made history when it became the first private company to land a spacecraft on the moon, so it makes sense to adapt that tech for Mars.

Intuitive Machines wants to help NASA return samples from Mars

As Google revamps itself for the AI era, offering AI overviews within its search results, the company is introducing a new way to filter for just text-based links. With the…

Google adds ‘Web’ search filter for showing old-school text links as AI rolls out

Blue Origin’s New Shepard rocket will take a crew to suborbital space for the first time in nearly two years later this month, the company announced on Tuesday.  The NS-25…

Blue Origin to resume crewed New Shepard launches on May 19

This will enable developers to use the on-device model to power their own AI features.

Google is building its Gemini Nano AI model into Chrome on the desktop

It ran 110 minutes, but Google managed to reference AI a whopping 121 times during Google I/O 2024 (by its own count). CEO Sundar Pichai referenced the figure to wrap…

Google mentioned ‘AI’ 120+ times during its I/O keynote

Firebase Genkit is an open source framework that enables developers to quickly build AI into new and existing applications.

Google launches Firebase Genkit, a new open source framework for building AI-powered apps

In the coming months, Google says it will open up the Gemini Nano model to more developers.

Patreon and Grammarly are already experimenting with Gemini Nano, says Google

As part of the update, Reddit also launched a dedicated AMA tab within the web post composer.

Reddit introduces new tools for ‘Ask Me Anything,’ its Q&A feature

Here are quick hits of the biggest news from the keynote as they are announced.

Google I/O 2024: Here’s everything Google just announced

LearnLM is already powering features across Google products, including in YouTube, Google’s Gemini apps, Google Search and Google Classroom.

LearnLM is Google’s new family of AI models for education

The official launch comes almost a year after YouTube began experimenting with AI-generated quizzes on its mobile app. 

Google is bringing AI-generated quizzes to academic videos on YouTube

Around 550 employees across autonomous vehicle company Motional have been laid off, according to information taken from WARN notice filings and sources at the company.  Earlier this week, TechCrunch reported…

Motional cut about 550 employees, around 40%, in recent restructuring, sources say

The keynote kicks off at 10 a.m. PT on Tuesday and will offer glimpses into the latest versions of Android, Wear OS and Android TV.

Google I/O 2024: Watch all of the AI, Android reveals

Google Play has a new discovery feature for apps, new ways to acquire users, updates to Play Points, and other enhancements to developer-facing tools.

Google Play preps a new full-screen app discovery feature and adds more developer tools

Soon, Android users will be able to drag and drop AI-generated images directly into their Gmail, Google Messages and other apps.

Gemini on Android becomes more capable and works with Gmail, Messages, YouTube and more

Veo can capture different visual and cinematic styles, including shots of landscapes and timelapses, and make edits and adjustments to already-generated footage.

Google Veo, a serious swing at AI-generated video, debuts at Google I/O 2024

In addition to the body of the emails themselves, the feature will also be able to analyze attachments, like PDFs.

Gemini comes to Gmail to summarize, draft emails, and more

The summaries are created based on Gemini’s analysis of insights from Google Maps’ community of more than 300 million contributors.

Google is bringing Gemini capabilities to Google Maps Platform

Google says that over 100,000 developers already tried the service.

Project IDX, Google’s next-gen IDE, is now in open beta

The system effectively listens for “conversation patterns commonly associated with scams” in-real time. 

Google will use Gemini to detect scams during calls

The standard Gemma models were only available in 2 billion and 7 billion parameter versions, making this quite a step up.

Google announces Gemma 2, a 27B-parameter version of its open model, launching in June

This is a great example of a company using generative AI to open its software to more users.

Google TalkBack will use Gemini to describe images for blind people

Google’s Circle to Search feature will now be able to solve more complex problems across psychics and math word problems. 

Circle to Search is now a better homework helper

People can now search using a video they upload combined with a text query to get an AI overview of the answers they need.

Google experiments with using video to search, thanks to Gemini AI

A search results page based on generative AI as its ranking mechanism will have wide-reaching consequences for online publishers.

Google will soon start using GenAI to organize some search results pages

Google has built a custom Gemini model for search to combine real-time information, Google’s ranking, long context and multimodal features.

Google is adding more AI to its search results