与其限制人工智能和算法，不如让它们变得可解释_AI阅读总结 — 包阅AI

The steady increase in deployment of AI tools has led a lot of people concerned about how software makes decisions that affect our lives. In one example, its about “algorithmic” feeds in social media that promote posts that drive engagement. A more serious impact can come from business decisions, such as how much premium to charge in car insurance. This can extend to affecting legal decisions, such as suggesting sentencing guidelines to judges.

Faced with these concerns, there is often a movement to restrict the use of algorithms, such as a recent activity in New York to restrict how social media networks generate feeds for children. Should we draw up more laws to fence in the rampaging algorithms?

In my view, the restricting the use of algorithms and AI here isn’t the right target. A regulation that says a social media company should forego its “algorithm” for a reverse-chronological feed misses the fact that a reverse-chronological feed is itself an algorithm. Software decision-making can lead to bad outcomes even without a hint of AI in the bits.

The general principle should be that decisions made by software must be explainable.

When a decision is made that affects my life, I need to understand what led to that decision. Perhaps the decision was based on incorrect information. Perhaps there is a logical flaw in the decision-making process that I need to question and escalate. I may need to better understand the decision process so that I can alter my actions to get better outcomes in the future.

A couple of years ago I rented a car from Avis. I returned the car to the same airport that I rented it from, yet was charged an additional one-way fee that was over 150% of the cost of the rental. Naturally I objected to this, but was just told that my appeal against the fee was denied, and the customer service agent was not able to explain the decision. As well as the time and annoyance this caused me, it also cost Avis my future custom. (And thanks to the intervention of American Express, they had to refund that fee anyway). That bad customer outcome was caused by opacity – refusing to explain their decision meant they weren’t able to realize they had made an error until they had probably incurred more costs than the fee itself. I suspect the error could be blamed on software, but probably too early for AI. The mechanism of the decision-making wasn’t the issue, the opacity was.

So if I’m looking to regulate social media feeds, rather than ban AI-driven algorithms, I would say that social media companies should be able to show the user why a post appears in their feed, and why it appears in the position it does. The reverse-chronological feed algorithm can do this quite trivially, any “more sophisticated” feed should be similarly explainable.

This, of course, is the rub for our AI systems. With explicit logic we can, at least in principle, explain a decision by examining the source code and relevant data. Such explanations are beyond most current AI tools. For me this is a reasonable rationale to restrict their usage, at least until developments to improve the explainability of AI bear fruit. (Such restrictions would, of course, happily incentivize the development of more explainable AI.)

This is not to say that we should have laws saying that all software decisions need detailed explanations. It would be excessive for me to demand a full pricing justification for every hotel room I want to book. But we should consider explainability as a vital principle when looking into disputes. If a friend of mine consistently sees different prices for the same goods, then we are in a position where justification is needed.

One consequence of this limitation is that AI can suggest options for a human to decide, but the human decider must be able to explain their reasoning irrespective of the computer suggestion. Computer prompting always introduces the the danger here that a person may just do what the computer says, but our principle should make clear that is not a justifiable response. (Indeed we should consider it as a smell for human to agree with computer suggestions too often.)

I’ve often felt that the best use of an opaque but effective AI model is as a tool to better understand a decision making process, possibly replacing it with more explicit logic. We’ve already seen expert players of go studying the computer’s play in order to improve their understanding of the game and thus their own strategies. Similar thinking uses AI to help understand tangled legacy systems. We rightly fear that AI may lead to more opaque decision making, but perhaps with the right incentives we can use AI tools as stepping stones to greater human knowledge.