What experts can learn by tracking AI harms
By Katie Peek
January 16, 2026
As artificial intelligence seeps more into people’s lives, making it ethical as well as functional is one research team’s goal. The group of academics, who hail from different institutions under the banner of the Responsible AI Collaborative, have been indexing stories of AI’s harmful outcomes since 2018. Back then, team leader and fellow at Harvard’s Berkman Klein Center Sean McGregor says, “we were peaking in terms of AI optimism without a balancing recognition of tradeoffs.” The researchers’ creation, the AI Incident Database (AIID), would help AI practitioners identify and address the technology’s weak points.
McGregor and his team compiled the first data points solely based on news stories. In 2020, they made the AI Incident Database public, and today anyone can submit an entry for consideration. Most incidents are anchored by a story in the press about, say, students creating deepfake pornography of classmates, or a wrongful arrest based on erroneous facial recognition. As a result, the database is not an exhaustive archive of all AI’s problems, but rather a compilation of its newsworthy issues. It captures emerging risks and especially significant issues in AI adoption.
The number of incidents in the AIID has increased over time, with fully half of the 861 entries appearing since 2022. The data is current as of December 2024. Thirty-four incidents that happened before 2014 are not pictured.
Incidents in the database are a mix of errors that come from AI’s development or implementation and events where AI tools are employed for nefarious purposes. They can be classified into a handful of broad categories:
Dots have white centers if a human being was harmed or killed.
AI behaving badly: systems with insufficient human oversight or otherwise needing better guardrails; includes inaccuracies, hallucinations and copyright infringements
AI on the move: events involving self-driving vehicles or autonomous robots
Bias: algorithms that perpetuate racial, gender and other biases
Content generation or moderation that is inappropriate, unethical or otherwise problematic; includes problems with social media algorithms
Fraud: deepfakes, LLMs and other Al tools used to defraud or scam
Privacy violations and surveillance, including facial recognition errors
Disinformation: deepfakes, LLMs and other Al tools used to spread false information
Other
Some categories, such as self-driving vehicles (in red), saw more activity in the past and are relatively quiet today; others, such as AI used for fraud (in light blue), are more recent problems that continue to grow. Here’s a look at how trends have developed since 2014:
2014 – 2016
AI on the move
July 2016
An autonomous security robot knocked down a 16-month-old toddler while patrolling a mall in Palo Alto, California. The robot then ran the child over, leaving him with swelling and a scrape.
Other
October 2014
Uber developed an AI tool to tag would-be users who were regulators, law officers and competitors in order to deny them rides. The program, called Greyball, helped Uber operate in cities where regulators had not yet approved it.
2018 – 2019
Bias
September 2018
Dutch authorities in Rotterdam used AI algorithms to flag residents to investigate for welfare fraud. The program showed baked-in biases based on age, gender and immigration status.
Content moderation
March 2019
In a report released following a mass shooting in Christchurch, New Zealand officials blamed YouTube’s recommendation algorithms for radicalizing the shooter.
2020 – 2021
AI behaving badly
April 2020
An algorithm designed to help television cameras track the soccer ball across the field mistakenly tracked a match official’s bald head instead.
Disinformation
October 2020
A climate action group in Belgium posted a deepfake video in which country’s premier appeared to state that the coronavirus pandemic was a byproduct of climate change.
2022 – 2023
AI behaving badly
May 2023
Tessa, an AI chatbot used by the National Eating Disorders Association, came under fire for giving some users advice — such as calorie restriction methods — that can exacerbate disordered eating.
Fraud
December 2023
One academic study cataloged large language models that are used by cybercriminals. The LLMs were particularly effective at malware code and phishing emails.
2024 – 2025
Privacy violations and surveillance
June 2024
The New York Times reported in June that both smartphone apps and auto manufacturers use AI to collect data on users’ driving habits, then make that information available to insurance companies.
Disinformation
January 2024
The athletic director of a high school near Baltimore used deepfake technology to create a clip of the school’s principal making racist and antisemitic statements about students.
AI behaving badly
April 2020
An algorithm designed to help television cameras track the soccer ball across the field mistakenly tracked a match official’s bald head instead.
AI behaving badly
May 2023
Tessa, an AI chatbot used by the National Eating Disorders Association, came under fire for giving some users advice — such as calorie restriction methods — that can exacerbate disordered eating.
AI on the move
July 2016
An autonomous security robot knocked down a 16-month-old toddler while patrolling a mall in Palo Alto, California. The robot then ran the child over, leaving him with swelling and a scrape.
Bias
September 2018
Dutch authorities in Rotterdam used AI algorithms to flag residents to investigate for welfare fraud. The program showed baked-in biases based on age, gender and immigration status.
Content moderation
March 2019
In a report released following a mass shooting in Christchurch, New Zealand officials blamed YouTube’s recommendation algorithms for radicalizing the shooter.
Fraud
December 2023
One academic study cataloged large language models that are used by cybercriminals. The LLMs were particularly effective at malware code and phishing emails.
Disinformation
October 2020
A climate action group in Belgium posted a deepfake video in which country’s premier appeared to state that the coronavirus pandemic was a byproduct of climate change.
Privacy violations and surveillance
June 2024
The New York Times reported in June that both smartphone apps and auto manufacturers use AI to collect data on users’ driving habits, then make that information available to insurance companies.
Other
October 2014
Uber developed an AI tool to tag would-be users who were regulators, law officers and competitors in order to deny them rides. The program, called Greyball, helped Uber operate in cities where regulators had not yet approved it.
Disinformation
January 2024
The athletic director of a high school near Baltimore used deepfake technology to create a clip of the school’s principal making racist and antisemitic statements about students.
Identifying problems, of course, is only the first step in solving them. But running the database has given the team unique insight into possible fixes. Humans should remain in the loop somewhere, McGregor says, rather than letting AI loose without oversight. Better guardrails and more careful training datasets can also help. Ultimately, the team hopes the AI Incident Database will help establish a culture of safety in the field, so that past errors won’t be repeated.
Katie Peek is a science journalist and data visualization designer based in Baltimore. Her work has appeared in Scientific American, where she ... Read More
Katie Peek
Generously supported by
the Future of Life Institute
Together, we make the world safer.
The Bulletin elevates expert voices above the noise. But as an independent nonprofit organization, our operations depend on the support of readers like you. Help us continue to deliver quality journalism that holds leaders accountable. Your support of our work at any level is important. In return, we promise our coverage will be understandable, influential, vigilant, solution-oriented, and fair-minded. Together we can make a difference.
Keywords: AI, LLM, artificial intelligence, bias, disinformation, harm, surveillance
Topics: Artificial Intelligence, Incoming Signals, The AI Power Trip
Recent Stories
Autonomous cars and trucks are a stupid idea. No computer can sense the world around them like a human can. Humans have brains with two halves, and both are useful in decision making. A human can sense that someone is about step off the curb into a street for example. A human can sense when someone is acting crazy or angry, can read and understand facial expressions and body language. A human driving a taxi or ride share car can help a passenger in many many ways that a computer can’t. A human taxi driver knows the city like only… Read more »
With the emphasis in AI on turning massive profits to justify the massive investments, I don’t see a culture of safety – or security – emerging any time soon. Cyberattackers already know patterns of bad AI coding practices and are using those in their exploits. Yes, humans have made mistakes as well, but the AI impact is to take bad practices and make them scale to the entire enterprise.