Technologies

AI Is Bad at Sudoku. It’s Even Worse at Showing Its Work

Researchers did more than ask chatbots to play games. They tested whether AI models could describe their thinking. The results were troubling.

Chatbots are genuinely impressive when you watch them do things they’re good at, like writing a basic email or creating weird, futuristic-looking images. But ask generative AI to solve one of those puzzles in the back of a newspaper, and things can quickly go off the rails.

That’s what researchers at the University of Colorado at Boulder found when they challenged 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).

A 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 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.

Don’t miss any of our unbiased tech content and lab-based reviews. Add CNET as a preferred Google source.

The paper is part of a growing body of research into the behavior of large language models. Other recent studies have found, for example, that models hallucinate in part because their training procedures incentivize them to produce results a user will like, rather than what is accurate, or that people who use LLMs to help them write essays are less likely to remember what they wrote. As gen AI becomes more and more a part of our daily lives, the implications of how this technology works and how we behave when using it become hugely important.

When you make a decision, you can try to justify it, or at least explain how you arrived at it. An AI model may not be able to accurately or transparently do the same. 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.

Read more: 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 — 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 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.”

Technologies

Anthropic warns investors of AI’s ‘existential risk to humanity’ in IPO prospectus, reports say

Anthropic’s IPO filing dedicates 80 pages to the risks of AI, nearly double the 48 pages it spends discussing its business plans, according to Reuters.

Your Privacy

On this service, we and our vendors use cookies and other tools (“Cookies”) to store and access information on your device, such as device identifiers, IP address, and your browser type. Your data may be used to save and communicate your privacy choices; ensure security, prevent fraud, and debug our products and services; personalize advertising and content; for advertising and content measurement; to conduct audience research and services development; so we can improve our services and develop new ones; to match and combine offline data with your online activity; and for social features. We may share this data with select vendors with your consent. You can adjust your choices any time through the “Cookie Preferences” link in the footer of relevant Versant sites or in-app settings.

We will also use other Cookies that are essential or related to our services, including for security and fraud prevention. To learn more about some of our vendors, see the IAB and Google Vendors under each purpose. Visit our Cookie Notice and Privacy Policy to learn more.

These Cookies and SDKs are required for Service functionality, including security and fraud prevention, and to enable any purchasing capabilities. You can set your browser to block these tracking technologies, but some parts of the site may not function properly.

The choices you make regarding the purposes and entities listed in this notice are saved and made available to those entities in the form of digital signals (such as a string of characters). This is necessary in order to enable both this service and those entities to respect such choices.

Your data can be used to monitor for and prevent unusual and possibly fraudulent activity (for example, regarding advertising, ad clicks by bots), and ensure systems and processes work properly and securely. It can also be used to correct any problems you, the publisher or the advertiser may encounter in the delivery of content and ads and in your interaction with them.

Certain information (like an IP address or device capabilities) is used to ensure the technical compatibility of the content or advertising, and to facilitate the transmission of the content or ad to your device.

Cookies, device or similar online identifiers (e.g. login-based identifiers, randomly assigned identifiers, network based identifiers) together with other information (e.g. browser type and information, language, screen size, supported technologies etc.) can be stored or read on your device to recognise it each time it connects to an app or to a website, for one or several of the purposes presented here.

– Use limited data to select advertising 791 partners can use this purpose Advertising presented to you on this service can be based on limited data, such as the website or app you are using, your non-precise location, your device type or which content you are (or have been) interacting with (for example, to limit the number of times an ad is presented to you).

– Create profiles for personalised advertising 637 partners can use this purpose Information about your activity on this service (such as forms you submit, content you look at) can be stored and combined with other information about you (for example, information from your previous activity on this service and other websites or apps) or similar users. This is then used to build or improve a profile about you (that might include possible interests and personal aspects). Your profile can be used (also later) to present advertising that appears more relevant based on your possible interests by this and other entities.

– Use profiles to select personalised advertising 640 partners can use this purpose Advertising presented to you on this service can be based on your advertising profiles, which can reflect your activity on this service or other websites or apps (like the forms you submit, content you look at), possible interests and personal aspects.

– Create profiles to personalise content 261 partners can use this purpose Information about your activity on this service (for instance, forms you submit, non-advertising content you look at) can be stored and combined with other information about you (such as your previous activity on this service or other websites or apps) or similar users. This is then used to build or improve a profile about you (which might for example include possible interests and personal aspects). Your profile can be used (also later) to present content that appears more relevant based on your possible interests, such as by adapting the order in which content is shown to you, so that it is even easier for you to find content that matches your interests.

– Use profiles to select personalised content 234 partners can use this purpose Content presented to you on this service can be based on your content personalisation profiles, which can reflect your activity on this or other services (for instance, the forms you submit, content you look at), possible interests and personal aspects. This can for example be used to adapt the order in which content is shown to you, so that it is even easier for you to find (non-advertising) content that matches your interests.

– Measure advertising performance 916 partners can use this purpose Information regarding which advertising is presented to you and how you interact with it can be used to determine how well an advert has worked for you or other users and whether the goals of the advertising were reached. For instance, whether you saw an ad, whether you clicked on it, whether it led you to buy a product or visit a website, etc. This is very helpful to understand the relevance of advertising campaigns.

– Measure content performance 405 partners can use this purpose Information regarding which content is presented to you and how you interact with it can be used to determine whether the (non-advertising) content e.g. reached its intended audience and matched your interests. For instance, whether you read an article, watch a video, listen to a podcast or look at a product description, how long you spent on this service and the web pages you visit etc. This is very helpful to understand the relevance of (non-advertising) content that is shown to you.

– Understand audiences through statistics or combinations of data from different sources 579 partners can use this purpose Reports can be generated based on the combination of data sets (like user profiles, statistics, market research, analytics data) regarding your interactions and those of other users with advertising or (non-advertising) content to identify common characteristics (for instance, to determine which target audiences are more receptive to an ad campaign or to certain contents).

– Develop and improve services 688 partners can use this purpose Information about your activity on this service, such as your interaction with ads or content, can be very helpful to improve products and services and to build new products and services based on user interactions, the type of audience, etc. This specific purpose does not include the development or improvement of user profiles and identifiers.

– Use limited data to select content 180 partners can use this purpose Content presented to you on this service can be based on limited data, such as the website or app you are using, your non-precise location, your device type, or which content you are (or have been) interacting with (for example, to limit the number of times a video or an article is presented to you).

These Cookies and SDKs are used to collect data about your browsing habits, use of the Services, your preferences, and your interaction with advertisements across platforms and devices for the purpose of delivering targeted advertising content, both on our Services and on third party sites. Third-party sites and services also use Targeting Cookies to deliver content, including advertisements relevant to your interests on the Services. If you reject these Cookies or SDKs, you will see less relevant advertising.

Data collected under this category through Cookies and SDKs can also be used to select and deliver personalized content, such as news articles and videos.

Technologies

U.S., Iran hold separate mediator talks as Mideast oil exports hit war-time high

Iranian and U.S. officials reportedly met separately with mediators Monday to push for a ceasefire, as Middle Eastern crude exports rebounded to war-time high.

Iranian and U.S. officials reportedly met separately with mediators Monday to push for a ceasefire.

Middle Eastern crude exports rebounded this month to around their highest level since the war began in February.

An outright export ban by Washington on diesel would keep roughly 1.2 million barrels a day at home, and risks overwhelming Gulf Coast storage, according to Kpler.

The building of the destroyed Shajareh Tayebeh primary school is pictured surrounded by rubble in Minab in Iran’s Hormozgan province on August 10, 2026. Six months on, a team of AFP journalists obtained rare access to the school which has remained largely in ruins, its collapsed walls and twisted steel now surrounded by portraits of the children killed as Iranian flags flutter above the rubble. Iran says the attack on the school involved two separate US Tomahawk missiles but neither the United States nor Israel has officially claimed responsibility for the strike. (Photo by ATTA KENARE / AFP via Getty Images) /

Atta Kenare | Afp | Getty Images

U.S. and Iranian officials have reportedly held separate indirect talks with mediators on Monday, reviving efforts to end seven months of war as Tehran presses for a response to a revised ceasefire framework and Middle East crude exports climb to their highest level since the conflict began.

Iranian Foreign Minister Abbas Araghchi met Qatari mediators in New York, where he remained after the United Nations General Assembly, and reportedly said he expects Washington’s reply by Tuesday. “We talked about ideas and how to find solutions to fulfil Iran’s conditions,” Araghchi said, adding that he planned to return to Tehran once a response arrives. “Whenever the Qataris have a response, they know how to get it to us.”

Iran’s proposal, first presented on the sidelines of last week’s General Assembly, calls for a four- to five-day U.S. timeline to release frozen Iranian funds, lift sanctions on Iranian oil and end the naval blockade of Iranian ports, with nuclear talks beginning within seven days. Tehran has tied any reopening of the Strait of Hormuz to those steps.

President Donald Trump had on Sunday called the plan “unacceptable” and rejected it, telling reporters Iran wants a swift deal because of the economic pressure it is under. Speaking at the White House Monday, Trump said U.S. officials had spoken separately with mediators but gave no further details, adding: “We’re going to win. It’s going to go pretty quickly.”

The diplomatic push comes as physical supply data shows the war’s toll on the oil market easing. Middle Eastern crude exports rebounded this month to around their highest level since the war began in February, Kpler data showed. “Middle East crude exports at just under 80% of pre-conflict levels,” the tracking firm said in a note on Monday.

The Strait of Hormuz itself remains far from normal. Kpler’s real-time tracking showed total clearance through the strait at 10,591 kilobarrels a day on Saturday, against a pre-war baseline of 17,133.

Technologies

Oil prices fall as crude exports recover at Saudi Arabia’s Red Sea ports

Saudi Arabia has restored flows through the pipeline to around 3.5 million barrels per day, according to media reports.

Your Privacy

Data collected under this category through Cookies and SDKs can also be used to select and deliver personalized content, such as news articles and videos.