Many organizations have published AI principles this year. The EU released its ethics guidelines for trustworthy AI in April. In May, the OECD countries, including Canada, adopted their own set, and the G20 endorsed them in June. Montreal had already issued its declaration for responsible AI at the end of last year.
These documents focus on fairness, transparency, accountability and human oversight. Their publication is a positive step, but it raises the question of whom they are written for.
The organizations now beginning to use AI in practice are not only governments and large technology companies; they include small business owners. A shop testing an automated filter for job applications, a tour operator letting an ad platform choose its audience, a clinic installing a bot on its Facebook page. Few of them will read an OECD document. This article therefore translates the main principles into practical terms.
How does bias enter AI systems?
The clearest example is a story Reuters reported last October. Amazon built an experimental tool to rank job applicants, trained on résumés the company had received over about ten years. Most of those résumés came from men, and the system learned to prefer them. Amazon abandoned the project.
No one programmed the system to discriminate. It learned from historical data, and that data was unbalanced.
This is the key point about bias in AI: the machine reproduces the pattern it receives, efficiently and without question. If the pattern is unfair, the output is unfair as well.
Facebook encountered the same issue in March. It reached a settlement with civil rights groups in the US and agreed to stop allowing housing, employment and credit ads to be targeted by age, gender or zip code. A few days later, the US housing department nevertheless filed charges against it over how its ad system determines who sees which ads.
For a small business, the effect is usually less visible. Consider an ad audience built from past customers, known as a "lookalike" audience. If those customers were mostly from one age group or one country, the platform will find more people with exactly those characteristics. That may be appropriate, or it may mean the business never reaches the customers it wants to attract next. In either case, it is important to be aware of it.
What responsibilities come with customer data?
Many of these tools rely on customer data: email lists, purchase histories, chat logs. The simplest approach is to upload everything and let the tool process it.
A simpler rule than any policy applies: do not provide a tool with anything you would not be comfortable explaining to the customer directly. If you would hesitate to say "I uploaded your details to this service so it could predict what you will buy next," reconsider the decision.
This is not only a matter of courtesy. Europe's GDPR has been in force since May last year, and it applies to how you handle the data of people in the EU, even if your business is located elsewhere in the world. Many dive guests are European. The GDPR also gives people rights regarding decisions made about them solely through automated processing. This article does not constitute legal advice, but if your list includes European customers, an hour of reading on the subject is worthwhile.
Why should customers be told when a machine is involved?
The most debated example is Google Duplex, demonstrated at Google I/O last year. An AI called a hair salon to book an appointment, complete with hesitations and verbal acknowledgements, and the person on the other end did not realize it was a machine. The audience responded enthusiastically, but many people later expressed discomfort. Google subsequently stated that the system would identify itself.
That discomfort was a sound response. People generally do not object to interacting with a machine; they object to being led to believe it is a person.
For a business, the solution is simple and costs nothing. If the chat window is a bot, state it clearly. If a machine drafted a reply and a person reviewed it, that is acceptable. A bot should not be presented as "Julie from customer service." Customers forgive a great deal, but feeling deceived is not among those things.
Which decisions should remain with a person?
This is the most important rule.
If you can't explain a decision to a customer, don't let a machine make it.
Refusing someone service, charging one person a different price than another, deciding who receives a job interview, closing an account: these decisions require a person who can review the case and explain the reasoning. The machine can prepare the file, but it should not make the final decision.
Many decisions carry little risk, such as choosing between two subject lines or deciding which photo appears first in an ad. These can safely be delegated.
The difficulty lies in distinguishing honestly between the two categories, because a convenient tool tends to extend gradually into decisions that matter.
One practice helps. Before activating any tool that makes decisions about customers, ask the vendor two questions: Can I see why it made a given decision? Can a person on my team override it? If the answer to either is no, or is unclear, keep the tool away from important decisions. The business, not the vendor, will be accountable to the customer.
Why will AI ethics become a small business responsibility?
The OECD AI Principles are brief and worth twenty minutes of reading. They are written for governments, but the principles apply equally well to businesses.
At present, the ethics discussion takes place mainly among large companies, researchers and governments, which is where AI is currently deployed. These tools, however, continue to become less expensive and easier to use. Small business owners will eventually have powerful capabilities available in a single click, with no one in the organization whose role is to ask these questions.
At that point, the owner effectively becomes the ethics department. We recommend that businesses begin considering these questions now, while the stakes are low and while seeking practical AI consulting or a second opinion remains a measured discussion rather than a response to a problem.