Emojis, spaces and small tweaks to fake content can beat Facebook’s AI controls

Fake claims on numerous Facebook posts weren’t slapped with warning labels by Meta’s AI controls – here’s how they got passed.

Meta flooded Australians with more than 150 dodgy and dangerous posts but got off scot-free, it can be revealed.

The posts, including “nonsense conspiracy theories”, ran on Facebook without “fake content” warning labels from late last year and were presented to the Digital Industry Group Inc (DIGI) for investigation.

DIGI is a non-profit industry association advocating for the digital industry in Australia. However this means it wears two hats – not only administering the voluntary Australian Code of Practice on Disinformation and Misinformation but also representing the tech giants themselves.

Online safety group Reset Tech Australia (RTA), which lodged the complaint claimed the posts had been fact-checked posts by a third party and found to be false.

RTA gave examples where small changes to wording were enough to beat Facebook’s automated controls, including extra spaces, capital letters and emojis.

Meta’s automated checks on misinformation have come under the spotlight. Picture: AFP

RTA said Meta’s 2023 Transparency Report was potentially misleading because it stated it applied “a warning label to content found to be false by third-party fact checking organisations”.

But DIGI’s independent subcommittee found there was no “convincing evidence” the transparency report was misleading and Meta didn’t have to correct the record – although the social media behemoth said it would clarify its “moderation and fact-checking process” in the next report

Dr Rob Nicholls, a policy and regulatory specialist at the University of Sydney, said there was a limit to DIGI’s independence.

Reset Australia Policy Director Rys Farthing and Executive Director Alice Dawkins. Picture: Benjamin Horgan.

“That’s because the complaints sub-committee, who are excellent people, are also members of the [Digi] administration sub committee and there’s no real reason to do that.

“It lacks rigour in an environment that really requires rigour,” Dr Nicholls said.

Reset Tech Australia Executive Director Alice Dawkins said the decision showed Australia’s self regulatory model had left the country “under protected’.

“Without public transparency and proper accountability, this is largely a ‘he said she said’ exercise where the onus is on the complainant to produce evidence that is virtually impossible to produce,” Ms Dawkins said.

She said the result was an “embarrassment” for Australia that once “led the world in online safety”.

The Reset investigation found “nonsense conspiracy theories” on Facebook were being posted with warning labels, including one with the question: “Did you know ONLY 2 countries are still considered Sovereign? Russia & Australia.”

It was found to be false by Meta’s fact-checkers and labelled with a warning, along with other identical or “near-identical” posts.

Dr Rob Nicholls.

But when it was tweaked to read: “There’s only two countries that are still sovereign in the world. Australia and Russia.” it, and others like it, were left unlabelled.

Ms Dawkins said Meta needed to “publicly acknowledge the limitations of (the) labelling system”.

“It creates a public impression that the majority of fact checked content is labelled by Meta – and crucially the remaining unlabelled content is accurate and trustworthy,” she said.

Meta denied its Transparency Report contained false statements and said it applied warning labels to content that is identical or near-identical to fact-checked content.

“This claim was dismissed as there was no evidence to suggest that Meta’s transparency report contained false information,” a Meta spokeswoman said.

Originally published as Emojis, spaces and small tweaks to fake content can beat Facebook’s AI controls