Grok, is that Gaza? AI picture checks mislocate information images

This picture by AFP photojournalist Omar al-Qattaa reveals a skeletal, underfed woman in Gaza, the place Israel’s blockade has fuelled fears of mass famine within the Palestinian territory.

But when social media customers requested Grok the place it got here from, X boss Elon Musk’s synthetic intelligence chatbot was sure that the {photograph} was taken in Yemen almost seven years in the past.

The AI bot’s unfaithful response was broadly shared on-line and a left-wing pro-Palestinian French lawmaker, Aymeric Caron, was accused of peddling disinformation on the Israel-Hamas battle for posting the picture.

At a time when web customers are turning to AI to confirm photographs increasingly, the furore reveals the dangers of trusting instruments like Grok, when the know-how is way from error-free.

Grok mentioned the picture confirmed Amal Hussain, a seven-year-old Yemeni little one, in October 2018.

In truth the picture reveals nine-year-old Mariam Dawwas within the arms of her mom Modallala in Gaza City on August 2, 2025.

Before the battle, sparked by Hamas’s October 7, 2023 assault on Israel, Mariam weighed 25 kilograms, her mom instructed AFP.

Today, she weighs solely 9. The solely vitamin she will get to assist her situation is milk, Modallala instructed AFP–and even that is “not always available”.

Challenged on its incorrect response, Grok mentioned: “I do not spread fake news; I base my answers on verified sources.”

The chatbot ultimately issued a response that recognised the error — however in reply to additional queries the subsequent day, Grok repeated its declare that the picture was from Yemen.

The chatbot has beforehand issued content material that praised Nazi chief Adolf Hitler and that recommended individuals with Jewish surnames had been extra more likely to unfold on-line hate.

– Radical proper bias –

Grok’s errors illustrate the boundaries of AI instruments, whose capabilities are as impenetrable as “black boxes”, mentioned Louis de Diesbach, a researcher in technological ethics.

“We don’t know exactly why they give this or that reply, nor how they prioritise their sources,” mentioned Diesbach, creator of a e book on AI instruments, “Hello ChatGPT”.

Each AI has biases linked to the data it was educated on and the directions of its creators, he mentioned.

In the researcher’s view Grok, made by Musk’s xAI start-up, reveals “highly pronounced biases which are highly aligned with the ideology” of the South African billionaire, a former confidante of US President Donald Trump and a standard-bearer for the unconventional proper.

Asking a chatbot to pinpoint a photograph’s origin takes it out of its correct function, mentioned Diesbach.

“Typically, when you look for the origin of an image, it might say: ‘This photo could have been taken in Yemen, could have been taken in Gaza, could have been taken in pretty much any country where there is famine’.”

AI doesn’t essentially search accuracy — “that’s not the goal,” the professional mentioned.

Another AFP {photograph} of a ravenous Gazan little one by al-Qattaa, taken in July 2025, had already been wrongly positioned and dated by Grok to Yemen, 2016.

That error led to web customers accusing the French newspaper Liberation, which had revealed the picture, of manipulation.

– ‘Friendly pathological liar’ –

An AI’s bias is linked to the information it’s fed and what occurs throughout fine-tuning — the so-called alignment part — which then determines what the mannequin would fee as or dangerous reply.

“Just because you explain to it that the answer’s wrong doesn’t mean it will then give a different one,” Diesbach mentioned.

“Its training data has not changed and neither has its alignment.”

Grok just isn’t alone in wrongly figuring out photographs.

When AFP requested Mistral AI’s Le Chat — which is partially educated on AFP’s articles below an settlement between the French start-up and the information company — the bot additionally misidentified the picture of Mariam Dawwas as being from Yemen.

For Diesbach, chatbots mustn’t ever be used as instruments to confirm details.

“They are not made to tell the truth,” however to “generate content, whether true or false”, he mentioned.

“You have to look at it like a friendly pathological liar — it may not always lie, but it always could.”

dou-aor/sbk/rlp