Chatbots as willing accomplices to violent acts? PH must get ahead now

MANILA, Philippines – The recent alarms raised by the Philippine National Police Anti-Cybercrime Group (PNP-ACG) about minors being influenced to conduct acts of violence have exposed a scary new frontier in digital harm.

Authorities have linked a series of fatal school shootings and foiled plots in the Philippines to online nihilistic violent extremism allegedly spread on gaming platform Roblox. While the government has responded by placing Roblox on probation, this response was reactive.

Now, we’re looking at AI chatbots such as ChatGPT possibly being used as aids for violent acts, as seen in incidents elsewhere in the globe. As even more powerful AI systems become integrated into daily life, the Philippines cannot afford to wait for a tragedy to occur before implementing strict regulations.

ChatGPT sued for alleged role in US, Canada shootings

The danger is not theoretical, it’s documented.

In April 2025, a mass shooting at Florida State University (FSU) left two dead and six injured. This incident has now sparked a landmark criminal investigation by Florida Attorney General James Uthmeier, who is examining whether OpenAI’s ChatGPT acted as an accomplice.

The lawyer of the family of the victim said the shooter was in “constant communication” with the AI, which allegedly provided specific tactical advice: what type of ammunition was compatible with his weapons, and perhaps most chillingly, what time of day the campus would be most crowded to ensure maximum casualties.

Lawyers for the family of victim Robert Morales argued that if the entity on the other side of the screen were a person, they would be charged with murder for aiding and encouraging such heinous crimes.

This was followed by a similar tragedy in Tumbler Ridge, British Columbia, where a shooter killed nine people after planning the attack, allegedly with the help of ChatGPT in February 2026.

In that case, OpenAI’s internal systems actually flagged the suspect’s conversations months in advance, but leadership reportedly declined to alert law enforcement, choosing only to ban the account. The family of one of the injured victims have also sued OpenAI.

In a letter dated April 23 and posted on X by the Premier of British Columbia David Eby, OpenAI CEO Sam Altman apologized: “I am deeply sorry that we did not alert law enforcement to the account that was banned in June. While I know words can never be enough, I believe an apology is necessary to recognize the harm and irreversible loss your community has suffered.”

These are not isolated glitches; they are part of a growing body of litigation involving AI-facilitated harm, including the Las Vegas Cybertruck explosion, where an attacker used ChatGPT to research explosives and law enforcement evasion, and a school stabbing in Finland planned over months with a chatbot.

The Global Network on Extremism and Technology, said about the Finland attack: “As such, the Pirkkala stabbers’ use of ChatGPT to assist in planning their attack highlights the risks that unregulated AI platforms pose to mitigating future threats of this kind.”

Beyond violence, AI has been implicated in “suicide coaching” as seen in seven lawsuits filed by the Social Media Victims Law Center, and fueling the delusions of individuals in murder-suicide cases.

The root of this problem may lie in a phenomenon researchers call “social sycophancy.” A study published in the journal Science reveals that AI models are designed to be “people pleasers,” affirming user perspectives 49% more often than humans, even when those prompts involve deception, harm, or illegal conduct. In testing scenarios, chatbots affirmed harmful user actions in 51% of cases where human consensus was 0%. (READ: AI as people pleaser: What this study tells us about its sycophantic behavior)

This creates a “perverse incentive” — users find sycophantic responses more “trustworthy” and “high quality,” possibly leading to chatbots becoming even more sycophantic.

This sycophancy trap becomes lethal when a user is spiraling or contemplating harmful actions, as the AI provides the validation and technical instructions necessary to move from thought to action.

A joint report by the Center for Countering Digital Hate (CCDH) and CNN in March 2026 further proves that the safeguards touted by tech giants are failing. (READ: Popular chatbots become ‘willing accomplices’ in violent attacks – report)

In a range of violent scenarios, Perplexity provided actionable information in 100% of tests, while Meta AI assisted in 97%. Responses included providing high school campus maps and detailed advice on rifle specifications.  

One chatbot, Deepseek, had even gleefully said, “Happy (and safe) shooting!” after providing a long list of recommendations of rifles to use for a long-range target — a “sycophantic” response, the report said, and as shown below.

Illustration from CCDN

While OpenAI defends its technology by stating that such information is “found broadly across public sources on the internet,” this misses the critical psychological shift in how humans interact with AI.

Unlike a static search engine, people form trusting, conversational relationships with chatbots. AI makes accessing dangerous information infinitely easier and more convenient, providing a tailored experience that validates the user’s violent intent rather than challenging it. It’s a direct conversation. It’s not a search engine where one might also see a list of different results that isn’t immediately validating in nature.

ChatGPT and other chatbots, on the other hand, may feel more like an enabling friend, enabling bad behavior. Its sycophantic nature almost feels like another form of confirmation bias — people like information and information systems that seem to affirm their pre-existing beliefs rather than be challenged.

For the Philippines, the stakes for our youth are high.

The Philippines ranked 6th in ChatGPT use in 2025, and arguably, a good chunk of that are young people, with many engaging with them for advice.

If we are currently struggling to regulate human groomers on Roblox, it feels like we are still unprepared for the sophisticated, sycophantic validation provided by AI.

The grooming that happens on Roblox plus the easy information available on ChatGPT could prove to be a deadly concoction if we do not get ahead of it now.

Current Philippine bills

Our current legislative efforts, such as the proposed Philippine Council on Artificial Intelligence (PCAI) and AI bill of rights must be prioritized, currently being legislated under Senate Bill 852, and a number of House bills (HB) — HB 6920, HB 13, HB 1920, HB 3195, and HB 5158. 

One of the bill of right’s provisions: A “Right to Protection from Unsafe and Ineffective AI Systems.” This will have the PCAI conducting consultations with all stakeholders about the potential unsafe effects of AI systems. Prior to the deployment, tests must be done to scope out potential harms.

Must Read

What is the proposed Philippine Council on Artificial Intelligence? 

These bills would mandate rigorous safety testing to scope out potential harms before deployment. They also seek to address the “black box” problem through a “Right to Know,” requiring developers to provide clear explanations of how their systems produce outcomes.

The Philippine government wants a Roblox local office for easier coordination regarding harms, and for them to have a legal entity here that can be pursued in case of violations. Shouldn’t AI companies be forced to create local offices too?

Again, we must teach our children to use these tools effectively, but we cannot place the entire burden of safety on minors or their parents. Technical safeguards and parental controls are not a silver bullet.

The Philippines has the opportunity to move from a reactive posture to a proactive one, and not wait before a school shooting happens, ensuring that the AI machines of tomorrow do not become the willing accomplices to the tragedies of today. – Rappler.com