There are always lines being drawn
I almost softened the “always” to be a “likely to be”, but decided not to. This post is about drawing lines: what’s in vs out, boundaries, policies around access and usage, and that even if we try to avoid it, there’s always a line to be drawn. Drawing lines can be offensive, because they involve judgement calls and difficult conversations. But there’s always lines — it’s a matter of where it’s drawn and who’s drawing it.
Where lines are drawn
I’ll start with some examples:
Example 1 — Advertising: Google/Facebook have clear policies on prohibited products, misrepresentation, copyright etc. Someone had to define and draw the line on prohibited goods, overselling, copying products, account for different regulatory nuances, public reaction etc.
Example 2 — Content moderation: A whole rabbit hole, but seeing the Fact Check section on Google News reminded me of this. Someone also had to define and draw lines on what are facts vs not.
Example 3 — Payment systems: Someone — usually central networks defines the lines of which end merchants are allowed to participate. There’s multiple parties. Regulators say no criminal activities. Payment networks may say “no controversial or high risk activities”. Banks may say “none of all of the above, and none of [additional ABC industries]” (even if legal).
Example 4 — Fundraising: Should a fundraising platform care about the causes that are listed on its platform? So even if they could process payments for all political parties, should they block some? Let’s say there’s a cause that you personally strongly care for, but 90% of people disagree strongly with you. Should you have access to raise funds for that cause from the 10%? And in the reverse, a cause you absolutely despise — but some chunk people want to fund raise for that. Is that OK? Will you let your opponents have the same access to fund raising?
Side note on (2) and (3) — there’s a correlation there given access to financial systems are at the heart of this. I think that’s what makes P2P networks so powerful — the only lines drawn are by individuals deciding if to pay another peer, assuming the payment network is not able to detect what the transaction is for 😅. But the facilitator of the p2p marketplace also gets to draw lines, so there’s no way to escape that.
I’ve given four examples of where tech has had to grapple with tough policy decisions. There’s more decision/line drawing ahead, in the buzzword field of AI.
Looking ahead — where lines will need to be drawn
I was initially reluctant to add to the already far-too crowded space on AI commentary. But I’m putting on the history major hat thinking through this. I believe lines are already being drawn, will continue to be drawn, and that we must pay attention to those lines.
Inputs and outputs
The quality of output depends on the inputs. Is all data of the same quality veracity? No. What happens if complete untruths are perpetuated and used to keep training the model? It’s often said that history is written by the victors — so who gets to write about what happens, or say what didn’t happened, and then gets to perpetrate untruths?
Judgement calls
I like how ChatGPT explains in its responses that it trains on “data created by human trainers”. I asked it to explain and paste an excerpt below. Bolded parts where I believe lines are drawn.
This training data is curated and prepared by human trainers who follow specific guidelines provided by OpenAI.
The trainers engage in a process called “unsupervised learning,” where they don’t provide explicit labels or corrections for every possible input. Instead, they generate and rank possible responses based on their own judgment and expertise. The training process involves exposing the model to a diverse range of text from the internet, books, articles, and other sources.
The human trainers review and curate this data, selecting examples that help improve the model’s understanding and response generation capabilities. They also follow guidelines to ensure that the model aligns with OpenAI’s values and policies, avoiding biased or inappropriate behavior.
- Specific guidelines, values, policies , curation — these are all words related to formally drawing lines.
- Exposing the model to a diverse range of text …— these are words that are more informal / judgement calls. While we can probably agree if something is NOT diverse, how wide is “diverse” here? What about fringe views? Who defines “diverse”? [1]
I ask those questions not because I have all the answers, but because I’m relieved that there are lines being drawn through guidelines, values, policies, curations, etc. We should pay attention because behind those guidelines, values, policies, there are human trainers drawing those lines. To no fault of their own, these are humans exposed to the joys, drudgeries and complexities — therefore experiencing the biases of life as they make these judgements. Whether we like it or not, there are always lines being drawn, shaped by humans.
[1] ChatGPT’s answer — Defining “diverse” in this context typically means including text from a variety of sources, such as different cultures, languages, regions, and subject matters. It aims to avoid over-representation or under-representation of specific viewpoints, demographics, or types of content. The goal is to create a model that can handle a wide range of topics and provide balanced and informative responses.
