AI, MISINFORMATION AND THE SHIFTING GROUNDS BENEATH OUR FEET.
My fellow Alums Shashank and Mayank had a small exchange about Disinformation etc. I made, what we called back then, a Class Participation point. I disappointed myself. To atone, let me give the well-intentioned conversation the respect it deserves. A quick essay at the cost of messing up my deadline.
__________________________________________________
We first encountered this issue in 2023; Digiday noticed it first.
I remember thinking that Dynamic Creative Optimisers and Automated CMS finally had a baby, with Programmatic acting as the wet nurse;
Right around then, AI broke out of IBM Watson/ Google Tensor Flow in-vitro environments. Unprecedented growth (in human history) and unimagined (Gartner hype cycle missed it) velocity – it was everywhere. For the first time in human history, the common people did not need a shaman to interpret the awesome new thing.
Cutting the nose to spite the face.
While the Real time, procedurally generative capabilities stupefied the whole fucking advertising and marketing scene – (profits plus products minus people); It excited the ones out to hustle an easy buck. We did not notice the slow burn havoc being germinated in the measurement, veracity and ethics side of things
Open web had Made for Advertising websites generated Realtime based on a search click-through, and vanishing right after they delivered the intended impression or click. Impossible to have a proper semantic scoring or a permanent record of these sites. Fine. We wrote papers on it.
The real shit fountain was bubbling within walled gardens, Particularly X. legions of trolls had access to the (free) GROKs seamless generation and propagation. A Clusterfuck of High Velocity, High Volume and highly vicious AI content.
GARM (Global Alliance for Responsible Media), advertiser-led industry initiative for improving brand safety and reduce harmful content, kept penalizing Twitter for this but my company kept giving X a safety rating of 99.9%
Elon gets punchy, Sues GARM.
No probs. Marketers and agencies passed the hat around.
Cue 2024 & Trump 2.0;
The US House of representatives Judiciary committee, among everything it could look at, decided for some reason to over-index its time on this. Publishes a 40-page report completely brutalizing GARM (and all associated clients and agencies). Poor clients who had been volunteering their time and energy got subpoenaed for phones, emails etc. Advertising’s happy little circle-jerk had the long grabby hands of the MAGA Jihad to contend with. The court documents are public. Ugly reading.
The large media agencies withheld spending on X for months. Couldn’t keep GARM Alive.
GARM, for all its faults and misplaced self-worth, was the only warden we had. We did not replace it with something better; We replaced it with a vacuum. Nature abhors a vacuum. Something will fill it for sure. Not fake-news watchdogs. They needed nurturing. META assumed that responsibility then rejected it. The lunatics have taken over the asylum.
BTW, You can read the House of Reps 40-page paper from 2024 here, good commute time pulp: Link
Wait there’s more; GARM is dead but still a convenient strawman. the 2024 angle was ONLINE Free SPEECH; The same government committee release a report a year later how GARM allowed foreign elements to silence American Patriotic voices: LINK
Here’s my thesis.
We are moving from the LLM past to the Self-RAG future (Self-Reflective Retrieval-Augmented Generation). AI is learning about itself. What we provide it as source is only a part of what it learns from. What we talk about it, What we write about it AND what we write about us is all part of the learning set.
If there’s bias in the ETHOS, there will be bias in the LOGOS.
Consumers, Let’s call them People, are (1) losing trust in brands and Ads; and (2) already getting the uncanny valley heebie-jeebies in otherwise normal, brand-safe, bog-standard encounters.
Remember the term Self RAG 80 odd words ago? Here’s why I mentioned it.
AI news summaries have errors; 51% of AI responses contained significant errors. Google Gemini fared the worst, with a60% error rate, while Perplexity performed the best (40%), per Ars Technica. Accuracy failures were the most common issue. Over 30% of responses significantly misrepresented facts, numbers, or dates. One in five AI-generated news summaries falsely attributed statements to BBC sources.
Alright mistakes happen. LLM.
13% of direct quotes were either altered or non-existent in cited BBC articles.
Editorializing was also a concern. AI summaries sometimes added bias absent from the original reporting. Added BIAS. Self-RAG.
We killed GARM and We also starved to destitution the good people who formed well-meaning organisations to do what they could. I will go back to the title of this essay.
QUIS CUSTODIET IPSOS CUSTODES