At a glance
- Company
- SiemensIndustrial manufacturer; the study was run at its electronics factory in Zug
- Industry
- Manufacturing
- Where and when
- Switzerland. June – July 2021, published December 2024.
- Who produced the figures
- Measured by researchersThe figures come from a preregistered experiment by university researchers, published in a peer-reviewed journal. One of the authors reports research funding from Siemens.
What they did
Workers at the Siemens factory in Zug check printed circuit boards by eye. They look for missing, wrong or faulty components. Researchers copied this job into an experiment and ran it on site with the factory’s own staff.
An AI system gave each board a quality score from 0 to 100. A low score meant a likely defect. Workers were told that the AI could be wrong. Each worker then decided whether the board was faultless or defective.
The workers were split at random into two groups:
- Score only. These workers saw the AI’s score and nothing else.
- Score plus a picture. These workers also saw the image with coloured areas marking where the AI had found a problem. The underlying AI was the same.
In all, 48 workers each judged 200 images of four types of board. Of the 200 boards, 28 were defective. The workers had 13.8 years of experience on average.
The numbers
| Defective boards caught | Overall accuracy | |
|---|---|---|
| Workers with the score only | 82.0% | 88.6% |
| Workers with the score and the picture | 93.0% | 96.3% |
| The AI alone | 92.9% | 95.6% |
Overall accuracy here gives equal weight to faultless and defective boards.
- People with the score alone did worse than the AI alone. They were 3.6 times more likely to overrule the AI when it was right.
- People with the picture did better than the AI alone. 73.1% of these workers beat it.
- They also caught its mistakes. Workers with the picture overruled 96.9% of the AI’s wrong predictions. Workers with the score alone overruled 86.4%.
What the case shows
- Adding a person to an AI can make the result worse. Experienced inspectors with a bare score threw away correct answers.
- How the AI presents its answer matters as much as how good it is. The same AI gave a better result once workers could see what it was reacting to.
- Test the pairing on the real task. The factory learned this from 48 people and 200 images in ten days.
- When you buy an inspection system, ask what the operator will see. A score with no reason attached is hard to judge.
What this says about the practices
Practices the company applied
- For each kind of task, test whether a person with AI beats either one alone The same inspection task was run with two kinds of AI aid and compared with the AI alone. Workers given only a score did worse than the AI alone. Workers shown where it had looked did better. Do next · Effort: medium (weeks, a few people)
- Keep a person checking anything that goes to a customer or carries legal weight A worker made the final call on every board. Workers who saw the AI’s reasoning overruled 96.9% of its wrong predictions. Do first · Effort: medium (weeks, a few people)
- Measure time saved per task against a baseline Every worker judged the same 200 product images under the same time limits, and the answers were scored against the known correct result. Do first · Effort: low (days, one person)
- Make the AI show what its answer is based on Half the workers saw the AI’s score with a picture marking where it had found a problem. They were more accurate than the workers who saw the score alone, and most beat the AI alone. Do next · Effort: medium (weeks, a few people)
Sources
- Senoner, Schallmoser, Kratzwald, Feuerriegel and Netland (2024). Explainable AI improves task performance in human–AI collaboration. Scientific Reports, 14, 31150. doi.org
- Senoner, Schallmoser, Kratzwald, Feuerriegel and Netland (2024). Explainable AI improves task performance in human–AI collaboration. Scientific Reports, 14, 31150. What this study did and found
- Vaccaro, Almaatouq and Malone (2024). When combinations of humans and AI are useful: A systematic review and meta-analysis. Nature Human Behaviour, 8(12), 2293–2303. What this study did and found