We will soon get used to using AI tools to help solve everyday problems and tasks. We must get in the habit of questioning motivations, motivations, and the capabilities behind them, too.
Imagine you are using an AI chatbot to plan a vacation. Did he suggest a particular resort because he knows your preferences, or because the company gets you bribe From the hotel chain? Later, when you use another AI chatbot to learn about a complex economic issue, does the chatbot reflect your policy or the policies of the company that trained it?
For AI to truly be our helper, it must be trustworthy. To be trustworthy, it must be within our control; Can not work behind the scenes to some technology monopoly. This means, at the very least, that technology needs to be transparent. And we should all understand how it works, at least a little bit.
Amidst the countless warnings about Suspicious Risks to well-being and threats democracyand even my presence The agony that has accompanied the amazing recent developments in artificial intelligence (AI) – and large language models (LLMs) such as conversationgpt And gPT-4– Very clear optimistic vision: This technology is useful. It can help you find information, express your ideas, correct errors in your writing, and much more. If we can overcome the pitfalls, then its auxiliary usefulness to mankind can determine the era. But we’re not there yet.
Let’s pause for a moment and imagine the possibilities of a reliable AI assistant. He can write the first draft of anything: emails, reports, articles, and even… wedding vows. You’ll have to give it background information and edit its output, of course, but this draft will be written by a model trained in your personal beliefs, knowledge, and style. It can act as your tutor, interactively answering questions about topics you want to learn about – in the way that works best for you taking into account what you already know. It can help you plan, organize, and communicate: again, depending on your personal preferences. It can defend you with third parties: either humans or other bots. It can moderate conversations on social media for you, flag misinformation, remove hate or trolling, translate for speakers of different languages, and keep discussions on topic; or even mediate conversations in physical spaces, interacting through speech recognition and synthesis capabilities.
Existing AI systems are not ready for this task. The problem isn’t technology – it’s advancing faster than ever Guess the experts– Who owns it. Existing AI systems are primarily created and managed by big tech companies, for their own benefit and profit. Sometimes we are allowed to interact with chatbots, but they don’t really belong to us. This is a conflict of interest, a conflict that destroys trust.
Going from terrible and eager use to skepticism to disappointment is reckless in the tech sector. Twenty years ago, it was the Google search engine quickly It rose to monopolistic dominance because of its ability to retrieve transformative information. Over time, the company’s reliance on revenue from search advertising led to this degrade that ability. Today, many observers look forward Until the search form is completely dead. Amazon has followed the same path, from a fair market to an over-the-top market shoddy products sellers who have paid To make the company show you. We can do better than this. If each of us is going to have an AI assistant who helps us with basic activities on a daily basis and even advocates on our behalf, then all of us need to know that he has our interests in mind. Building a trustworthy AI will require systemic change.
First, the trustworthy AI system must be controlled by the user. This means that the model must be able to run on user-owned electronic devices (possibly in simplified form) or within a cloud service they control. It must show the user how it responds to them, such as when it makes queries to search the web or external services, or when it instructs other programs to do things like send email on behalf of the user, or modify user data. summonedTo better express what the company that made it thinks the user wants. They must be able to explain their reasons to users and cite their sources. These requirements are all within the technical capabilities of AI systems.
Furthermore, users must control the data used to train and fine-tune the AI system. When modern LLMs are built, they are first trained on public body Text data usually obtained from the Internet. Many systems are progressing further than that fine tuning on more specific datasets that are designed for a narrow application, like speaking in terms of a file doctoror tradition Method and style of the individual user. In the near future, corporate AI systems will be routinely fed your data, possibly without your awareness or consent. Any trustworthy AI system must transparently allow users to control the data it uses.
Many of us would welcome an AI-assisted typing app fine-tuned knowing what edits we’ve accepted in the past and which we haven’t. We would be more skeptical of a chatbot knowing which of its search results led to purchases and which did not.
You should also be aware of what the AI system can do for you. Can he access other apps on your phone and the data stored in them? Can it retrieve information from external sources, mixing your input with details from other places that you may or may not trust? Can he send a message with your name (hopefully based on your input)? Balancing these kinds of risks and benefits will become an integral part of our daily lives as AI aids integrate into everything we do.
Realistically, we should all prepare for a world where AI is not very trustworthy. As AI tools can be so useful, they will increasingly pervade our lives, whether we trust them or not. Being a digital native in the next quarter of the 21st century will require learning the basic ins and outs of an LLM so you can assess its risks and limitations for a given use case. This will better prepare you to take advantage of, rather than take advantage of, AI tools.
In the first few months of widespread use of models like ChatGPT in the world, we have learned a lot about how artificial intelligence creates risks for users. By now everyone has heard that LLMs are “hallucinating”, meaning that they constitute “facts” in their output, because their predictive text generation systems are not limited to verifying the fact of their emission. many users I learned In March, information they send as prompts to systems like ChatGPT may no longer be kept secret after an error exposed users’ conversations. Your chat logs are stored in potentially insecure systems.
The researchers found several Maher Methods to trick chatbots into breaching their safety controls; These work largely because many of the “rules” that apply to these systems are soft, eg directions It is given to a person, rather than hard restrictions, such as coded restrictions on product functionality. It’s as if we’re trying to keep the AI safe by asking them to drive with careful, hopeful instructions, rather than pulling out its keys and setting specific limits on its capabilities.
These risks will only increase as companies give chatbot systems more capabilities. OpenAI provides developers at scale Access To create tools above GPT: tools that give their AI systems access to your email, to your personal account information on websites, and to your computer code. While OpenAI applies security protocols to these integrations, it’s not hard to imagine those relaxing the drive to make the tools more useful. It also seems inevitable that other companies will come up with less sneaky strategies to secure their AI market share.
Just as with any human being, building trust with AI will be challenging through interaction over time. We will need to test these systems in different contexts, observe their behavior, and build a mental model of how they respond to our actions. Trust can only be built in this way if these systems are transparent about their capabilities, what inputs they use and when they share them, and the interests of whom they evolve to represent.
Want to learn more about artificial intelligence, chatbots, and the future of machine learning? Check out our full coverage of artificial intelligenceor browse our guides to The best free AI art generators And Everything we know about OpenAI’s ChatGPT.
Nathan E. Sanders is a data scientist and affiliate member of the Berkman Klein Center at Harvard University.
Bruce Schneier is a security technologist and public policy lecturer at Harvard Kennedy School.