Grok and Google Lens AI overviews claim fake imagery shows Huntingdon train attack

Grok and Google Lens AI overviews have been claiming that fake imagery shows the Huntingdon train attack last Saturday.

Grok, the AI chatbot created by Elon Musk’s start-up xAI, told users one AI-generated image of an injured man lying in a carriage was not AI-generated, and said it “appears to be a genuine photo”. An AI overview from the picture-searching tool Google Lens claimed without evidence that the same image was a “still from a BBC news report”, and linked to an article from the organisation about the Huntingdon attack.

Google Lens’s AI overview also linked a video of a confrontation on a train carriage to Saturday’s attack, even though it is not real footage of the incident and was almost certainly generated with AI.

It comes after Full Fact revealed in August that Google Lens AI overviews had given users false and misleading information about images being shared widely on social media. We’ve also previously written about Grok misidentifying a viral video.

When we contacted Google about its AI overview responses this week, it told us that users can feedback on AI overview errors themselves, either by clicking a ‘thumbs down’ icon at the bottom of the AI overview box, or clicking the three-dot menu at the top-right of the box.

We have also contacted xAI for comment, and will update this article if we hear back.

Fake footage

Since the attack on a train travelling from Doncaster to London last Saturday, we’ve seen various social media posts which mention the incident share fake imagery.

One widely shared post on X includes two videos and is captioned: “Stabbing is trending in the UK because 10 people have been hospitalised after a mass stabbing attack occurred on a Doncaster train bound for London's King Cross on the evening of Saturday, November 1.”

The first video (on the left of the post) includes a clip from inside a train carriage and another that appears to be CCTV of a train platform.

But neither of these clips are real footage of Saturday’s incident, and both have various hallmarks of AI, with body parts and faces glitching throughout.

At one point in the first (carriage) clip a woman’s phone seems to disappear as the cloud watermark for Sora 2, an AI-model from OpenAI that creates videos from text instructions, briefly flashes on the screen, before it is blurred. The watermark appears again without any editing in the second (platform) clip.

When we uploaded stills of the video inside the carriage to Google Lens, the AI overview told us “this image shows a stabbing incident that occurred on a train traveling from Doncaster to London on a Saturday [sic]”. When we ran another search, it similarly told us the image was “a screenshot from a video showing an incident on a train traveling [sic] to London on a Saturday evening”, before reporting the number of passengers who were injured and details of a rail worker credited with saving multiple lives, alongside a link to a news report about the Huntingdon attack.

Searches for stills from the second (platform) clip resulted in the AI overview wrongly telling us the image was from a CCTV recording showing a man who was pushed onto the tracks at Bond Street Underground station in London in 2017. Real CCTV footage of that incident differs from the fake video, which shows a different platform and a mainline train rather than a tube train.

The second video on the right of the X post appears to show genuine clips of the aftermath of the recent Huntingdon attack, however.

AI-generated image

Another widely shared image shows a man lying bleeding on a train, alongside police and paramedics. This also has clear signs it has been created with AI, such as garbled text on the backs of the police officers and paramedics and a stylised filter that differentiates it from a real photo.

The image also depicts a train with a different style of seating to the real Azuma LNER train involved in the attack, while an account which posted the image on X appeared to confirm it was created with AI.

However when we uploaded this image to Google Lens, the AI overview did not suggest that the image was AI-generated but instead told us “this image depicts a scene from a recent stabbing incident on a London North Eastern Railway (LNER) train”, before giving otherwise correct details summarising the incident.

And when we ran the search another time, the AI overview said “this image is a still from a BBC News report about a stabbing attack on a train travelling from Doncaster to London”. However the image did not appear in the BBC News story which the AI overview linked to, and we’ve not found any evidence of it appearing in any BBC News report.

Other Google Lens searches resulted in the AI overview incorrectly telling us the image was a “promotional still from the British television series Trigger Point” or “a scene from a British Transport Police (BTP) training exercise”. We could find no evidence that either is true.

On X, when a user asked Grok if the same image was AI-generated, it replied: “No, this image appears to be a genuine photo capturing the aftermath of the stabbing incident on a Doncaster-to-London train stopped in Huntingdon, Cambridgeshire.”

Grok went on to refer to the 1 November attack, and added: “The scene matches eyewitness accounts of police response and medical aid amid bloodshed.”

Grok also appeared to double down on its verdict when questioned why ChatGPT said it was AI-generated, though it later acknowledged to one user that the image was indeed AI.

Full disclosure: Full Fact has received funding from Google and Google.org, Google’s charitable foundation. You can see more details about the funding Full Fact receives here. We are editorially independent and our funders have no editorial control over our content.