The year the world noticed AI — and started worrying about it

In 2023, 67% of respondents said they had a good understanding of what AI is, up from 64% a year earlier. 66% expected AI products and services to profoundly change their daily life in the next three to five years, up from 60%. 54% said those products have more benefits than drawbacks, up from 52%. And 52% said AI products and services make them nervous — up from 39%. That 13-point jump is more than twice the size of any other movement in the survey.

The most revealing pair of numbers is the one about time. 49% said AI products and services have already profoundly changed their daily life over the past three to five years — exactly the figure recorded in 2022 — against the 66% who expect that change in the next three to five. Nothing changed in the answer about the past; everything changed in the answer about the future. Only 51% say they know which products and services actually use AI, one point up on the year. Trust sat flat: 52% trust companies that use AI as much as they trust other companies (50% in 2022), 56% trust AI not to discriminate or show bias toward any group, and 50% trust that companies using AI will protect their personal data.

Nervousness rose in every single country

Of the 24 countries surveyed in both 2022 and 2023, every one reported an increase in the share saying AI products make them nervous — not one exception. The sharpest rises were Italy (24 points), France (19), Chile (18) and Australia (18); the mildest were South Africa (1), Japan (3) and Spain (3). Belief that AI will significantly change daily life within three to five years rose almost as broadly, with South Africa the only country to fall (down 2 points) and gains of 12 points each in Japan, Great Britain, Germany and Australia.

Where sentiment was lowest, it improved the most

In 2022 a cluster of developed Western nations — Germany, the Netherlands, Australia, Belgium, Canada and the United States — sat among the least positive about AI products and services. By 2023 every one of them had risen, and the Netherlands moved furthest: 43% of Dutch respondents saw AI as net beneficial, up from 33% a year earlier. Japan matched that 10-point gain to reach 52%. The direction was not universal, though. Chile fell 4 points, Peru and Spain 3 each, and Hungary, Poland and Sweden 1 each.

Two countries mark the ends of the range

Indonesia is the outlier at the optimistic end: 84% claim a solid understanding of AI, 79% expect it to significantly change their lives within three to five years, 78% see more benefits than drawbacks, and 75% say AI products make them excited. Japan is the outlier in the other direction and in an unexpected way — it reports the lowest understanding of AI of any country surveyed (43%) and, at the same time, by far the lowest nervousness (23%), against 63% in the United States. Thai respondents showed the highest trust in AI's impartiality, with 83% believing it will not discriminate or show bias toward any group. Among the 31 countries, the United States and France were the most skeptical about AI's benefits, both at 37%.

The 2023 Ipsos survey added questions about work. 57% said it is likely that AI will change how they perform their current job within the next five years, against 35% who said it is not likely and 8% who did not know. On outright replacement the answer inverts: 36% said it is likely AI will replace their current job in the same window, 56% said it is not, and 8% did not know. The public expects to keep the job and do it differently.

Who expects the change depends heavily on who you ask. 66% of Gen Z respondents agreed AI will likely affect how they do their current job, against 61% of millennials, 53% of Gen X and 46% of baby boomers — a 20-point generational spread. The other splits point the same way: 68% of those in decision-making roles agreed against 51% of those who are not, 61% of the highest-educated group against 52% of the lowest, and 60% of upper-income respondents against 56% of both middle- and lower-income respondents. Gender barely registers, at 58% for men and 55% for women.

Convenience, yes. Prosperity, no.

• 54% of global respondents agree AI will reduce the time it takes to get things done, and 51% expect better entertainment options — television, video, music and books. Those are the only two majorities in the whole set.
• Every question that touches livelihood falls below half: 39% think AI will improve their health, 37% their own job, 34% their national economy, and 32% the job market.
• The country pattern from the products-and-services questions repeats almost exactly. Japanese, Swedish and American respondents are consistently pessimistic about AI improving their lives; Brazilians, Indonesians and Mexicans are consistently optimistic.
• On health, Mexico leads at 61% against a global average of 39%, while Japan comes last at 16%. On whether AI will improve their own job, Thailand tops the table at 66% and Japan again sits at the bottom, at 19%.
• The job market draws the bleakest readings anywhere in the survey: 17% in South Korea, 18% in Belgium and 19% in Canada, against 54% in Mexico and Peru.
• Optimism tracks age, income and education together. 59% of Gen Z believe AI will improve their entertainment options against 40% of baby boomers, and on all six questions Gen Z, higher earners and the more educated score highest while boomers, lower earners and the less educated score lowest.

In 2021, 37% of Americans said they felt more concerned than excited about the increased use of AI in daily life. In 2022, 38%. By 2023, 52%. The move did not come out of the enthusiastic camp, which was already small and shrinking: those more excited than concerned fell from 18% to 15% to 10%. It came out of the middle. The share reporting themselves equally concerned and excited held at 45% and 46% before dropping to 36% in a single year.

Americans separate what AI does from what it knows about them

Pew also asked whether AI helps or hurts in eight specific settings. Only one drew a clear positive: helping people find products and services they are interested in online, at 49% helps against 15% hurts. Doctors providing quality care to patients and companies making safe cars and trucks both drew 37% helps. Two settings came out negative — companies providing quality customer service (28% helps against 34% hurts) and police maintaining public safety (24% against 26%). And one was not close at all: on people keeping their personal information private, 53% said AI hurts more than it helps against just 10% who said it helps. Uncertainty runs through the whole set, with 35% to 49% answering 'not sure' on every question.

Education changes the answer more than anything else

• On AI helping people find products and services online, 60% of Americans with a college degree or more said it helps, against 44% of those with some college or less.
• On doctors providing quality care, 46% of college graduates said AI helps against 32% of the less-educated group; on companies making safe cars and trucks, 45% against 33%; on people taking care of their health, 41% against 29%.
• The gap holds on customer service (34% against 25%) and on finding accurate information online (36% against 31%).
• It nearly disappears on policing. 27% of college graduates say AI helps police maintain public safety against 23% of everyone else — and 49% of both groups answer 'not sure'.
• Privacy is the exception that runs the other way: the more educated group is more negative, not less. 59% of college graduates said AI hurts people keeping their personal information private, against 50% of those with some college or less.

GPT-4 took 53% of all AI-related social media conversation in the first quarter of 2023, 71% in the second after its March 14 launch, 62% in the third and 45% in the fourth. That last figure is the striking one. Q4 2023 brought Gemini, Grok, GPT-4 Turbo, Claude 2.1, Whisper V3, GraphCast, Orca 2, Inflection-2, Midjourney v6 and Stable Video Diffusion — and nearly half of everything said about AI online was still about a model released in March. Of the models that launched in that quarter, only Grok (16%) and Gemini (11%) registered at all; every other one came in at 2% or below.

Stable Diffusion is the only other model that ever held a large share, and its arc runs the other way: 46% of conversation in Q1, then 21%, 24% and 12%. Copilot never exceeded 1% of AI conversation in any quarter of the year.

The best-liked models were the ones nobody was arguing about

Attention and approval moved in opposite directions all year. The highest net sentiment score of 2023 went to GraphCast, DeepMind's weather forecasting system, at 94 in Q4 — with a rounded 0% share of conversation. Claude 2.1 scored 87 on a 2% share; Whisper V3 and Orca 2 both scored 83, Inflection-2 81. GPT-4, the model everyone was talking about, never scored above 51 and dipped to 42 in Q1 and 44 in Q4. Gemini, launched into the most crowded quarter of the year, scored 36 — the lowest net sentiment on the chart.

What people actually argued about

• GPT-4's positive sentiment came from faster processing, better accuracy, and productivity gains in coding, corporate collaboration and content creation. The negative side came from ChatGPT outages and from the open letter led by Elon Musk and supported by more than 1,300 AI experts urging labs to pause training of powerful systems — a letter that was itself widely disputed, including by Andrew Ng.
• GPT-4 Turbo drew praise in Q4 for longer conversations, improved contextual understanding and image generation, and complaints about its April 2023 knowledge cutoff and slower loading than GPT-4.
• Stable Diffusion discussion concentrated in the first half of the year, and posts mentioned the XL models around 16 times more often than Stable Diffusion 2.0. Enthusiasm ran to creativity, resolution and accuracy; the negative half was copyright, ownership of AI-generated work, and the displacement of human artists.
• Gemini's Q4 conversation praised its accuracy, multilingual ability and potential inside Google Search and Ads, while criticism centred on inaccurate results, the delayed release, and skepticism toward the launch demo.