Oil rigs and giant harbours: PN chatbot hallucinates flawed images of proposals

The PN's innovative AI idea went rogue in less than 48 hours after it was released

Updated 3.45pm with PN reply

The Nationalist Party's manifesto chatbot has stopped generating certain images after users complained that it began producing unprompted, highly inaccurate visuals of party proposals, including an oil rig and a massive, unrealistic expansion of Gozo’s Mġarr harbour.

The abrupt technical change follows questions raised on social media by photographer and environmental activist Daniel Cilia, and comes just a day after the same chatbot caused some political embarrassment by contradicting the party’s official renewable energy targets.

Screenshots obtained by Times of Malta show that until Tuesday evening, the chatbot was automatically generating images to illustrate PN proposals, even when users did not request them. And some of them were inaccurate.

In a reaction on Wednesday, the PN said it altered its code to improve user experience and prevent further hallucinations.

The AI tool was only deployed to help voters navigate the party's proposals, a party spokesperson said, but "images are not architectural drawings, or technical renders and so do not represent the detail of particular proposals", he clarified.

A gigantic harbour and a rusty tanker

In one instance, a user asking in Maltese about plans for Gozo's Mġarr Harbour was shown a sprawling, oversized terminal featuring five large, comfortably docked ferries, an expansive marina and several apartment blocks. The visual clearly requires a land footprint that does not exist in the actual location.

The bot generated inaccurate images even when the user did not ask for pictures.The bot generated inaccurate images even when the user did not ask for pictures.

In a post on Facebook, Cilia posted the image and wrote: "This must be a joke or a scam! PN proposals for Imġarr Harbour!"

In a more glaring error, a user asked about the PN’s so-called offshore fuel station project. In its answer, the bot described the proposal fairly well, even specifying that the station would use clean energy.

But right after the description, the bot generated an image of a traditional oil rig refuelling a rusty tanker at sea, completely undermining the party's environmental messaging.

The bot correctly said the project would use clean fuels...The bot correctly said the project would use clean fuels...

...only to then generate an image of an oil rig refuelling a rusty tanker out at sea....only to then generate an image of an oil rig refuelling a rusty tanker out at sea.

Bot refusing to generate images

By Wednesday morning, the chatbot had stopped producing images, simply telling users it would only answer questions about the PN manifesto instead.

But some users had already posted the images on social media.

Writing on Facebook, engineer Simon Mizzi, who is among Alex Borg's closest aides and advisors, defended the party’s official Gozo manifesto, stressing that the official document makes "no reference to this visual".

He urged the public to separate reality from AI generation, saying everybody knows by now that this is what AI often does, and people should not interpret those images as accurate representations of what the party intends to do if elected to government.

"AI is a visual tool like anything else, and it gives you a visual image that is not always correct... I think it’s not the first time you've made your judgment on this - on what is truth and what is AI," he wrote.

The controversy is meant to confuse, rather than inform, he added.

'Strong engagement in first 36 hours' - PN

In a reply to questions the PN said the tool recorded "strong engagement" in the first 36 hours since it was deployed, and this "shows that people are actively using the tool to explore how PN’s proposals relate to their own lives".

"In some cases, illustrative images are generated automatically as part of the AI experience, even when users do not explicitly request them. These visuals are intended purely as conceptual aids, helping make policy content easier to understand for people who process information better through images, including some users with learning difficulties or those who may find long policy text harder to follow," he said.

"As with every AI tool, following internal review and feedback received, adjustments are constantly made to the experience, including restricting the automatic generation of some images. This is done to avoid potential confusion and to keep users focused on the actual policy content and explanations set out in the programme."

This is the second time in 48 hours that the PN's tech-driven campaign tool has gone rogue.

On Tuesday, energy minister Miriam Dalli flagged a conflict within the PN manifesto, incorrectly claiming the party had two clashing targets for renewable energy.

When prompted about it, the chatbot itself turned on its own party, admitting that there was, in fact, a conflict.

Sign up to our free newsletters

Get the best updates straight to your inbox:

You can unsubscribe at any time by clicking the link in the footer of our emails. We use Mailchimp as our marketing platform. By subscribing, you acknowledge that your information will be transferred to Mailchimp for processing.