The Survey and Notions Behind It

AI is on everyone’s lips. While John, sitting in the office cubicle, is considerably free to decide if to use ChatGPT or Microsoft Copilot in his working routine, business and social leaders are expected to keep tabs on the situation. And they do, expressing an enormously wide range of opinions towards “How AI will be used in the future” and the effects AI provides on society and human life.

On one part of the spectrum are people who consider AI an innovative technology capable of improving the quality of life in its various aspects. Bill Gates’ post written at the end of 2023 is a good example. As a Microsoft cofounder states, AI saves children's lives, helps reduce pregnancy risks, and assesses people’s risk for HIV – and AI solves this task unprecedentedly effectively.

On the other side of the scale are scientists, science fiction writers, and simply interested individuals discussing AI takeover – a hypothetical scenario where artificial intelligence becomes the dominant form of intelligence on Earth, taking control away from humans. It’s the most pessimistic vision of how AI will impact the future.

Meanwhile, a moderate attitude towards AI prevails. As said at World Economic Forum's Annual Meeting at Davos (January 2024), “AI can contribute to solving the biggest challenges of humanity; however, guardrails are vital”.

This balanced perspective recognizes the risks of AI, such as its potential impact on public opinion, widening socioeconomic gaps, misuse in surveillance, and the effects on employment - pointing to the need for strategic approaches that address safety through responsible guidelines and prevent adverse effects.

Image Description

Scientists contribute to the discussion actively, operating by solid figures. One of the most frequently cited in 2023 surveys, ESPAI, focuses on the future of AI, encompassing the anticipated timelines for achieving key AI milestones, the societal and ethical implications of AI advancements, and the challenges in AI safety and alignment. Since the survey is based on authors of NeurIPS and ICML papers opinion’s, it ensures a high level of credibility, drawing from a pool of contributors at the forefront of AI research and development.

One of the survey peculiarities is that it was conducted for the third time. It gave the scientists the chance to compare the results and get meaningful insights about how AI will be used in the future. The first two surveys were conducted in 2016 and 2022. The interesting fact is that how notions changed between “iterations” is more considerable between 2022 and 2023 compared to six years between two earlier research studies. It means that the AI field is transforming before our very eyes. Let's keep our finger on the pulse of these rapid changes by examining the scientists' findings in more detail.

How Soon Will AI Replace Human Professionals?

This question, allegedly, excites the imagination the most. To answer this question, researchers used the concept of High-Level Machine Intelligence (HLMI) – the state when machines not supported by people can handle ALL possible tasks more efficiently and cheaply compared to humans.

To assess the closeness of the AI apocalypse the prospect of technologies capable of outperforming humans in productivity, scientists asked respondents to evaluate 39 tasks that help to find out how AI will be used in the future. Regarding each task, respondents had to answer the question: "Within what timeframe will AI be able to perform this operation?". Here is the shortened list of evaluated tasks, including relatively simple and predictable to complex and sometimes unexpected ones:

• put together LEGO following guidelines;
• identify an object after a single viewing;
• arrange clothes neatly;
• create a video from a novel perspective;
• train a Large Language Model;
• develop a website for handling financial transactions;
• produce a new song in the style of a particular musician;
• compete in the World Series of Poker;
• write a quality essay on high school history.

It occurred that AI apo the day when AI is able to train itself and a robot can fold the laundry is much closer than we could think. Thirty-five tasks were estimated to be done by artificial intelligence within the decade. Only a few tasks were estimated as requiring more time to come to reality:

• install the electrical cables in a newly constructed house manually (17 years);
• conduct research and author documents (19 years);
• validate mathematical principles (22 years).

When will we achieve Human-Level Machine Intelligence (HLMI)? This question stands out as the most captivating aspect of the survey. Originally, the consensus in the 2022 survey pointed towards 2060 as the year we could expect HLMI to be realized. However, the 2023 survey presents a significant shift in expectations. Respondents now anticipate that HLMI could be achieved by 2047, marking a dramatic adjustment of thirteen years earlier than previously thought. It's interesting to note that prior adjustments in estimates were much more modest; for instance, the forecast only shifted by one year, from 2061 to 2060, between the 2016 and 2022 surveys.

Is the prospect of AI reaching a level where it significantly surpasses human capabilities in a wide range of tasks closer than we might anticipate?

What Will AI look like in 2047?

Let's lift the curtain on the future a bit more. We've learned what AI is predicted to be capable of doing in the future. But what will it be like? Researchers are interested in two questions: AI's vulnerabilities and its ability to evaluate its own decisions, that is, to look at itself from an outside perspective.

AI’s Vulnerabilities

In 2043, AI systems are expected to be quite advanced, as we found out. Nevertheless. experts are still unsure about how they'll turn out and how AI will impact the future. Study authors looked into whether future AI will have certain qualities like being trustworthy, able to make decisions on their own, and even breaking out of their programmed limits. Out of 11 key qualities, experts think it's pretty likely that AI will have most of them by 2043.

But, there are a few things most experts think AI will be able to do by 2043. A good number believe that AI will do the following things:

• find clever ways to meet goals;
• talk just like a human expert on many topics;
• often act in ways that surprise us.

This shows that even if there's a lot of uncertainty about how AI will be used in the future, we can expect it to be both highly capable and unpredictable in some fascinating ways.

AI’s Self-Explainability

Looking ahead, the big question is: will AI be able to clearly tell us why it did what it did? The feedback from a survey of experts isn't too hopeful. Most of them think it's unlikely that AI systems will be open books about their decision-making.

This issue isn't just about understanding AI for curiosity's sake. It's about making sure AI's decisions are fair, transparent, and accountable. If an AI system decides something important, like who gets a loan or how medical treatment is given, we need to know the "why" behind its decision. Unfortunately, as of now, it looks like we'll still be guessing at those reasons in 2028. This connects to a broader concern about whether future AI, even by 2043, will be trustworthy enough to accurately share what's going on under its hood. The overall vibe? We're facing a challenge in making AI as open and understandable as we'd like it to be.

How AI Will be Used in the Future: Wrapping Up

As we navigate through the evolving landscape of artificial intelligence, the "2023 Expert Survey on Progress in AI" offers both a hopeful outlook and cautionary insights into the future. It reveals an accelerating pace of development, suggesting that tasks once thought exclusively human could soon be within AI's grasp, with potential realizations of High-Level Machine Intelligence significantly sooner than previously anticipated. However, amidst this technological leap forward, the survey underscores critical areas of concern – particularly AI's vulnerabilities and its capacity for self-explanation. These findings highlight the importance of fostering AI systems that are not only advanced but also transparent, accountable, and aligned with ethical standards. As we stand on the brink of these advancements, the dialogue between optimism and careful scrutiny remains more crucial than ever, guiding us toward a future where AI enhances humanity without compromising our values or autonomy.