A practical list of embodied AI safety concerns: Polluting the information space (Part 4)
A somewhat different type of harmful content is the unintentional creation of misinformation, or the purposeful creation of disinformation. Photos of things that never happened and misleading LLM-generated material packaged as news reports are proliferating. Some outputs of chat systems are shared by uninformed people as if they were true.
Some false content is shared for entertainment, with the assumption that readers will recognize satire when they see it.1 Some false content is intentionally created as a weapon of propaganda. This is an inevitable outcome of cheap creation of synthetic images, video, and text by systems which do not have a concept of “truth” embedded in them. There do not seem to be any simple, robust ways to address this problem.
LLMs and other generative AI techniques are ideal for creating propaganda at scale, which can do significant societal harm. Especially problematic are disinformation campaigns that are not really intended to get recipients to deeply believe something specific, but rather to produce a flood of credible-sounding false information. The goal is not to fool people on any individual item, but rather to reach a point at which people become skeptical of anything they hear, degrading the effect of objectively true statements and scientifically rigorous findings.2
An alternate use of LLMs and other generative AI techniques is to generate coordinated disinformation to get people to believe something that is simply not true, in a technique known as the “big lie.”3 Automated generation of an endless stream of variations on the same falsehood can be accomplished inexpensively and at scale with this technology. One can go further and associate different synthetic personas with different social media accounts and generate statements of the big lie with those synthetic voices. And so on. Note that the LLMs are not actually lying here – they are spouting BS as usual. However, the prompts submitted by the disinformation campaigners set the scene to stipulate that the big lie is a fact, and generate BS but seemingly authentic content from there. Political propagandists never had it so good.
Even if there is no malicious intent, the at-scale creation of synthetic text and images pollutes the information space. As the Internet currently operates, posting more content can create more ad revenue, so there is an incentive to create as much content as possible, with authenticity being of little concern for what has become known as content farming.4 Synthetic content can be biased to increase engagement, disregarding truth, creating further revenue.
However, at some point the synthetic content overwhelms organic human-created content.5 This applies not only to Internet text, but also to visual illustrations and books, including some that are outright scams.6 Audiobooks and podcasts are already seeing AI-generated content, and more sophisticated synthetic videos are on the way. That makes it difficult for people to find content created by a real person, which was created with some notion of truth, or was created with the individual flair of a particular creator rather than a statistical mashup started by a prompt. These creations might be informative, entertaining, or otherwise useful to some audiences. But when they masquerade as or drown out human-created content, that becomes a problem.
Many issues can arise from a tsunami of synthetic content. The most obvious harm is to human creators who are displaced from the marketplace, sometimes with works that either imitate them or even fraudulently claim to have been created by them.
A more subtle harm is degrading people’s trust in anything they see on the Internet. This is especially problematic when mainstream sources of information use an AI/ML-based workflow in an attempt to cut costs and keep up with a content-creation arms race.
On social media, synthetic content can be used to automate placing disinformation into discussions, drowning out dissenting points of view with an onslaught of contrary postings and ad hominem attacks against the poster. One expects there are whole communities of echo chambers to indoctrinate those who find their way there into certain ways of thinking.
A more ironic problem with information pollution is that AI systems that train on the output from AI systems are prone to so-called model collapse.7 That is a phenomenon in which an AI/ML system trained on the outputs from other AI/ML systems (or itself) degrades instead of improving with additional training. As the Internet becomes choked with synthetic content, models will have trouble finding pristine training data. This seems likely to form a natural limit on the current race to build ever-bigger AI training data sets.
To the extent that embodied AI (eAI) devices will present information to users and influence their attention, they can be coopted into presenting disinformation.8 They might also be biased in their data presentation in ways that might be unintentional, or perhaps purposeful. As an example, consider an eAI pair of glasses that helps provide information about people surrounding a user. What if one of the features is a personal risk indicator that the user comes to rely on to estimate how cautious they should be around others in a public setting? Sounds like it might be useful for personal safety. But what if the eAI training data is biased to indicate an artificially high risk when the user is near some population demographic that is politically out of favor?
Or what if an automated public incident recording system is biased to artificially elevate the severity of a public order infraction? If a typical-looking person bumps into someone else due to not paying attention, that is not recorded. But if an out-of-favor minority member brushes against someone else in a packed public area, that is considered a physical assault and recorded as such (and perhaps biased data is later used to “prove” that demographic is prone to violence and should have restrictions placed upon them). Nothing good can come of such a system, but that type of abuse of AI/ML technology is entirely foreseeable.
As with creating harmful content, mitigation approaches have yet to be created. This will be a challenging area for some time to come.
Next posting: Additional eAI ethical issues
This post is a draft preview of a section of my new book that will be published in 2025.
In practice, this is a questionable assumption.
Having people believe the lies amounts to a bonus, but is not the primary goal of such campaigns. See Paul & Matthews, 2016: https://www.rand.org/pubs/perspectives/PE198.html
Arguably some parts of the Internet are already there, with one report claiming 31.4% of all on-line content, 60% of blog posts, and 80% of e-commerce product descriptions are AI-generated.
See: https://www.clrn.org/how-much-of-the-internet-is-ai-generated-content/
HOWEVER: The Grammarly AI detector claimed that 45% of the text from that posting was AI generated. And GPTZero claims 100% probability that the post was AI-generated. So really, from this we have no idea what the fraction of material on the Internet is generated by AI, because that blog page was probably just making the numbers up. Or the AI detectors (based on AI/ML technology) are lying to us. Who knows… Which proves our larger point here.
We think of it as mad cow disease for AI.
See: https://en.wikipedia.org/wiki/Model_collapse
While the incident was blamed on a supposed rogue employee, for about a day one LLM was modified to aggressively push a controversial racial genocide narrative. See Tufekci, 2025: https://www.nytimes.com/2025/05/17/opinion/grok-ai-musk-x-south-africa.html