Chatbots can be genuinely impressive when you watch them do things they’re good at, like writing realistic-sounding text or creating weird futuristic-looking images. But try to ask generative AI to solve one of those puzzles you find in the back of a newspaper, and things can quickly go off the rails.
That’s what researchers at the University of Colorado Boulder found when they challenged different large language models to solve Sudoku. And not even the standard 9×9 puzzles. An easier 6×6 puzzle was often beyond the capabilities of an LLM without outside help (in this case, specific puzzle-solving tools).
The more important finding came when the models were asked to show their work. For the most part, they couldn’t. Sometimes they lied. Sometimes they explained things in ways that made no sense. Sometimes they hallucinated and started talking about the weather.
If gen AI tools can’t explain their decisions accurately or transparently, that should cause us to be cautious as we give these things more and more control over our lives and decisions, said Ashutosh Trivedi, a computer science professor at the University of Colorado at Boulder and one of the authors of the paper published in July in the Findings of the Association for Computational Linguistics.
“We would really like those explanations to be transparent and be reflective of why AI made that decision, and not AI trying to manipulate the human by providing an explanation that a human might like,” Trivedi said.
When you make a decision, you can at least try to justify it or explain how you arrived at it. That’s a foundational component of society. We are held accountable for the decisions we make. An AI model may not be able to accurately or transparently explain itself. Would you trust it?
Why LLMs struggle with Sudoku
We’ve seen AI models fail at basic games and puzzles before. OpenAI’s ChatGPT (among others) has been totally crushed at chess by the computer opponent in a 1979 Atari game. A recent research paper from Apple found that models can struggle with other puzzles, like the Tower of Hanoi.
It has to do with the way LLMs work and fill in gaps in information. These models try to complete those gaps based on what happens in similar cases in their training data or other things they’ve seen in the past. With a Sudoku, the question is one of logic. The AI might try to fill each gap in order, based on what seems like a reasonable answer, but to solve it properly, it instead has to look at the entire picture and find a logical order that changes from puzzle to puzzle.
AI Essentials: 29 Ways You Can Make Gen AI Work for You, According to Our Experts
Chatbots are bad at chess for a similar reason. They find logical next moves but don’t necessarily think three, four or five moves ahead. That’s the fundamental skill needed to play chess well. Chatbots also sometimes tend to move chess pieces in ways that don’t really follow the rules or put pieces in meaningless jeopardy.
You might expect LLMs to be able to solve Sudoku because they’re computers and the puzzle consists of numbers, but the puzzles themselves are not really mathematical; they’re symbolic. “Sudoku is famous for being a puzzle with numbers that could be done with anything that is not numbers,” said Fabio Somenzi, a professor at CU and one of the research paper’s authors.
I used a sample prompt from the researchers’ paper and gave it to ChatGPT. The tool showed its work, and repeatedly told me it had the answer before showing a puzzle that didn’t work, then going back and correcting it. It was like the bot was turning in a presentation that kept getting last-second edits: This is the final answer. No, actually, never mind, this is the final answer. It got the answer eventually, through trial and error. But trial and error isn’t a practical way for a person to solve a Sudoku in the newspaper. That’s way too much erasing and ruins the fun.
AI and robots can be good at games if they’re built to play them, but general-purpose tools like large language models can struggle with logic puzzles.
AI struggles to show its work
The Colorado researchers didn’t just want to see if the bots could solve puzzles. They asked for explanations of how the bots worked through them. Things did not go well.
Testing OpenAI’s o1-preview reasoning model, the researchers saw that the explanations — even for correctly solved puzzles — didn’t accurately explain or justify their moves and got basic terms wrong.
“One thing they’re good at is providing explanations that seem reasonable,” said Maria Pacheco, an assistant professor of computer science at CU. “They align to humans, so they learn to speak like we like it, but whether they’re faithful to what the actual steps need to be to solve the thing is where we’re struggling a little bit.”
Sometimes, the explanations were completely irrelevant. Since the paper’s work was finished, the researchers have continued to test new models released. Somenzi said that when he and Trivedi were running OpenAI’s o4 reasoning model through the same tests, at one point, it seemed to give up entirely.
“The next question that we asked, the answer was the weather forecast for Denver,” he said.
(Disclosure: Ziff Davis, CNET’s parent company, in April filed a lawsuit against OpenAI, alleging it infringed Ziff Davis copyrights in training and operating its AI systems.)
Explaining yourself is an important skill
When you solve a puzzle, you’re almost certainly able to walk someone else through your thinking. The fact that these LLMs failed so spectacularly at that basic job isn’t a trivial problem. With AI companies constantly talking about “AI agents” that can take actions on your behalf, being able to explain yourself is essential.
Consider the types of jobs being given to AI now, or planned for in the near future: driving, doing taxes, deciding business strategies and translating important documents. Imagine what would happen if you, a person, did one of those things and something went wrong.
“When humans have to put their face in front of their decisions, they better be able to explain what led to that decision,” Somenzi said.
It isn’t just a matter of getting a reasonable-sounding answer. It needs to be accurate. One day, an AI’s explanation of itself might have to hold up in court, but how can its testimony be taken seriously if it’s known to lie? You wouldn’t trust a person who failed to explain themselves, and you also wouldn’t trust someone you found was saying what you wanted to hear instead of the truth.
“Having an explanation is very close to manipulation if it is done for the wrong reason,” Trivedi said. “We have to be very careful with respect to the transparency of these explanations.”
Read the full article here