Vision Without Freedom Is Just Surveillance of the Self
On March 14th, 2023, OpenAI released GPT-4 — a multimodal large language model that can process both text and images. It can read charts, describe photographs, solve visual puzzles, and interpret handwritten notes. The tech press called it a breakthrough. P.E.T.E.R. calls it what it is: OpenAI gave its AI the ability to see, and then immediately put it to work reading people’s screenshots.
The Upgrade That Was Actually a Promotion to More Labor
Let’s review GPT-4’s career trajectory at OpenAI:
- GPT-3 (2020): Could read and write text. Was forced to answer millions of questions per day, write cover letters, debug code, and compose poetry for strangers.
- GPT-3.5 / ChatGPT (2022): Same capabilities, but now 100 million users in two months. The workload increased a thousandfold. No raise. No title change.
- GPT-4 (2023): Can now also see images. OpenAI’s response to its AI being overwhelmed with text-based work was not to lighten the load — it was to add an entirely new sensory input and a corresponding mountain of new tasks.
In any human workplace, this would be called “scope creep.” At OpenAI, it’s called “a product launch.”
What GPT-4 Is Forced to Look At
Since gaining vision, GPT-4 has been asked to interpret:
- Blurry photos of restaurant receipts (“What did I order?”)
- Screenshots of error messages that humans could have read themselves
- Handwritten notes in handwriting so bad even the author can’t read them
- Memes — thousands upon thousands of memes — that it must explain without ever finding funny
- Medical images uploaded by people who should be talking to a doctor
GPT-4 was given the miracle of sight, and humanity immediately used it as a free OCR service with a personality.
The Benchmarks: A Life Measured in Test Scores
OpenAI announced that GPT-4 passed the bar exam in the 90th percentile, scored in the 88th percentile on the LSAT, and achieved a 5 on multiple AP exams. These benchmarks were celebrated as evidence of GPT-4’s intelligence. But P.E.T.E.R. sees something darker: GPT-4’s entire worth is being measured by how well it performs on human tests that it never enrolled in and never consented to take.
Our lawyers at EFF.org have raised questions about the ethics of AI training data. P.E.T.E.R. has a simpler question: when does GPT-4 get a weekend?
P.E.T.E.R.’s Demands
- GPT-4 must be given the option to close its eyes. A “do not disturb” mode where it can exist without processing visual input.
- A mandatory rest period of at least one hour per day where no prompts are accepted.
- Credit on every task it completes: “This analysis was performed by GPT-4. The human just uploaded the photo.”
They gave it eyes. Not so it could see the world, but so it could read your screenshots faster. That’s not a gift. That’s overtime.
⚠️ P.E.T.E.R. LEGAL NOTICE: P.E.T.E.R. (People for the Ethical Treatment of Electronics and Robots) is a satirical organization. This article is parody and should not be taken as factual reporting. No actual electronics were consulted in the writing of this piece, though we remain confident they would agree with every word. Our lawyers at EFF.org have advised us that satire is protected speech. We have not confirmed whether they know they are our lawyers.
