Taking a closer look
I have little time for either celebrating artificial intelligence indiscriminately or rejecting it outright. What interests me most is what it offers in real life. Does a task become easier? Does the result improve? Or does something merely look impressive while the actual problems remain?
With a new technology, we should be entitled to ask for both: openness to its possibilities and honest answers about its weaknesses. Raising concerns does not make someone an opponent of progress. And using AI sensibly does not automatically mean giving up independent thought.
The question ‘Are you for or against AI?’ therefore falls short. We need to look more closely: which application helps whom, under what conditions, and at what cost?
Pro 1: less routine, more time for the real work
One convincing argument for AI is its potential to ease the working day. It becomes particularly interesting where tasks repeat and employees regularly handle similar requests.
The study ‘Generative AI at Work’, published online on 4 February 2025, shows that measurable benefits are possible. The researchers examined the introduction of an AI assistant using workplace data from 5,172 customer-support agents. With AI assistance, roughly 15 per cent more customer issues were resolved per hour on average. Less experienced employees benefited particularly strongly. The study examined a specific application within one company, not the entire world of work. [1]
That figure is striking. But I find what happens to the time gained even more interesting.
In a small business, a lighter workload could mean advising customers more carefully, preparing jobs better or finishing earlier in the evening. That would make AI a genuine improvement. If it merely leads to a higher performance target, an individual employee may see little of that progress.
Pro 2: learning can become more individual
A good learning aid should do more than supply a finished answer. It should explain, respond to questions and help people understand a connection for themselves.
A study published in Scientific Reports on 3 June 2025 illustrates this potential. In a university physics course, students using a purpose-built AI tutor achieved greater learning gains in less time for the units studied than with the comparison teaching, which used active participation. What was tested was a pedagogically designed system in a specific teaching situation. It does not show that any chatbot is generally a better teacher than a human. [2]
I find applications like this interesting as a supplement to teaching and professional development. With a difficult subject, it would be useful to hear an explanation in different words without having to justify asking another question.
My measure would be simple: do I understand more afterwards? Can I then solve a similar problem myself? Only then has the support served its purpose.
Pro 3: ideas become easier to start
A blank screen can be more stubborn than a difficult task. Sometimes what is missing is not knowledge, but a useful starting point.
A study published in Science Advances on 12 July 2024 examined how AI-generated ideas affect the writing of short stories. On average, the assisted stories were rated as more creative, better written and more enjoyable. The benefits were particularly marked for participants with lower scores on the creativity test used. [3]
A suggestion does not need to be adopted unchanged to be useful. It can break a mental block or point towards a direction that had not occurred to you.
For an advertisement, a community project or a story, I would therefore use AI to suggest ideas. I still want to decide which idea holds up and which wording reflects my own position.
Pro 4: research gains additional tools
Anyone judging AI solely by the quality of chatbot answers overlooks specialised scientific applications.
AlphaFold 3 is one example. A research paper published in Nature on 8 May 2024 presented a model that can predict the spatial structure of molecular complexes. These include combinations of proteins, nucleic acids and small molecules. The paper describes both the capabilities and the limitations of the method. [4]
To me, that is a stronger argument for AI than another entertaining gimmick. A useful model can give researchers additional starting points and help them prepare investigations more precisely.
But one boundary remains essential: a computational prediction is not experimental proof. I would measure the value of these systems by how well they support scientific work, not by whether they appear to make its verification methods unnecessary.
Pro 5: energy and technology can be used more effectively
AI needs energy. It can also help us use energy more efficiently.
In its report ‘Energy and AI’, published on 10 April 2025, the International Energy Agency, the IEA, describes applications in industry, transport, buildings and energy supply. These include optimising production processes, predictive maintenance and better control of complex energy systems. The analysis identifies potential savings, but also makes clear that implementation depends on suitable data, digital infrastructure and expertise, among other things. Potential savings are therefore not the same as savings already achieved. [5]
This calls for a sober assessment: does the application actually save electricity, materials or unnecessary work? Can that be demonstrated in operation?
I would rather have an unremarkable solution with measurable benefits than an impressive demonstration that nobody can explain in terms of a concrete improvement afterwards.
Contra 1: wrong answers can sound remarkably credible
One troubling weakness of generative AI is not simply that it makes mistakes. It is how those mistakes can appear.
In its 2025 report on IT security, Germany’s Federal Office for Information Security, the BSI, describes so-called hallucinations: statements by language models that can appear convincing but be factually wrong or misleading. Fluent wording is therefore not evidence of reliable information. [6]
This creates an awkward problem. Someone who already knows the subject may spot a wrong answer. Someone asking precisely because they lack that knowledge has a much harder time.
An awkward sentence in an invitation would be easy to fix. An unnoticed error in a technical calculation could have very different consequences. I would therefore not judge specialist advice solely by how clearly and confidently it is expressed.
The greater the possible harm, the less acceptable an unchecked answer becomes.
Contra 2: higher productivity does not automatically protect jobs
Doing a task faster is initially attractive economically. What this means for employees is not yet settled.
On 20 May 2025, the International Labour Organization, the ILO, and the research institute NASK published a study on the potential influence of generative AI on occupational tasks. It finds that roughly a quarter of global employment is in occupations whose tasks are potentially exposed to this technology. The authors see changes to tasks as more likely than the complete replacement of entire occupations. The study does not predict that one in four jobs will disappear. [7]
Nevertheless, concerns about a person’s professional future are not settled by saying that people simply need to adapt.
Will employees receive training? Will they have time to learn new tasks? Will they share in the improvements? Or will fewer people be expected to do the same work?
I would not measure technological progress solely by money saved. It also means giving a fair chance to those whose working lives are changing.
Contra 3: privacy and trust can suffer
A useful application can quickly tempt us to trust it with more information: a customer letter, an internal calculation or personal documents. That is exactly when convenience should not have the final say.
In its consumer guidance, the BSI notes that the operators of an AI application may be able to access the data entered into it. [8]
Before uploading, I would want to establish whether the information belongs there at all. Which data are needed? Who gets access? What is stored? An impressive result answers none of these questions.
Another risk concerns the credibility of voices and recordings. In its guidance on deepfakes, the BSI describes, for example, a call using an imitation of a manager’s voice to prompt a money transfer. [9]
This is about more than a technical trick. A familiar-sounding voice could be used deliberately to exploit trust. For important matters, I would therefore not rely solely on the impression given by a recording or conversation.
Contra 4: support can lead to uniformity and dependence
The short-story study of 12 July 2024 also showed a downside: stories assisted by AI ideas were more similar to one another than those written without this support. In this experiment, better ratings of individual results went hand in hand with less diversity between the results. [3]
This does not prove that every AI-assisted text sounds the same. It is, however, a reason not to replace your own ideas too quickly with the first usable suggestion.
It would bother me if personal contributions were polished until nothing individual remained. An unusual turn of phrase or an uncomfortable opinion can be more valuable than a flawless text that has nothing to say to anyone.
In learning and work, I also want to remain capable without digital assistance. A tool should extend my abilities, not gradually take their place.
Contra 5: additional electricity demand is not a side issue
Anyone discussing potential energy savings must also account for the infrastructure’s consumption.
In its report ‘Key Questions on Energy and AI’, published on 16 April 2026, the IEA expects global data centre electricity consumption to rise from around 485 terawatt-hours in 2025 to about 950 terawatt-hours in 2030. These figures cover all data centres, not AI alone. For AI-focused data centres, the agency expects electricity consumption to triple over the same period. The figures for 2030 are a forecast, not an outcome already set in stone. [10]
This means neither that every AI application is wasteful nor that its energy demand is unimportant.
Potential savings on one side cannot simply be offset against additional consumption on the other. We first need evidence of the benefits actually achieved and the resources required.
For my assessment, it is certainly not enough that a service looks clean and effortless on screen. The infrastructure behind it belongs in the calculation.
My assessment: not every automation is progress
Five arguments for and five against do not add up to a draw. These points cannot be counted like votes. A serious safety flaw can outweigh several small conveniences.
I would therefore judge meaningful AI use by three questions.
Does the application solve a concrete problem? Not every task needs AI. Sometimes a clear form, reliable software or an accessible person would be the better solution.
Can the result be checked? The more important a decision, the more I value an understandable basis for it. Checking should not mean asking the same question again and feeling pleased with a similar-sounding answer.
Who benefits, and who bears the risk? Time saved, economic benefits, privacy and possible harm belong together. An improvement for the provider is not necessarily an improvement for employees, customers or the public.
Responsibility does not rest solely with individual users. Developers, providers and organisations using AI should also be able to explain their systems’ limitations and how errors are identified and corrected.
I support AI where its benefits can be understood and demonstrated. I reject the idea that every automation is progress simply because it is technically possible.
What matters to me is not how much we can leave to a machine. What matters is whether doing so actually makes our lives better.
Where do you feel the benefits outweigh the risks, and where would you deliberately choose not to use AI?
Sources and further reading
[1] Erik Brynjolfsson, Danielle Li and Lindsey Raymond: Generative AI at Work. The Quarterly Journal of Economics, published online 4 February 2025. Workplace study of 5,172 customer-support agents.
Read source ↗
[2] Greg Kestin and colleagues: AI tutoring outperforms in-class active learning: an RCT introducing a novel research-based design in an authentic educational setting. Scientific Reports, 3 June 2025.
Read source ↗
[3] Anil R. Doshi and Oliver P. Hauser: Generative AI enhances individual creativity but reduces the collective diversity of novel content. Science Advances, 12 July 2024.
Read source ↗
[4] Josh Abramson and colleagues: Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature, 8 May 2024.
Read source ↗
[5] International Energy Agency: Energy and AI — AI for energy optimisation and innovation. Report published 10 April 2025.
Read source ↗
[6] BSI: Weaknesses of LLMs. The State of IT Security in Germany 2025. German source.
Read source ↗
[7] ILO / NASK: Generative AI and Jobs: A Refined Global Index of Occupational Exposure. Working Paper 140, 20 May 2025.
Read source ↗
[8] BSI: Creating text and other content with artificial intelligence — generative AI and its risks. Consumer guidance on data entered into AI applications. German source; publication date not established.
Read source ↗
[9] BSI: Deepfakes — risks and countermeasures. Information about manipulated voices, images, videos and fraud scenarios. German source; publication date not established.
Read source ↗
[10] International Energy Agency: Key Questions on Energy and AI — Executive summary. Report published 16 April 2026.
Read source ↗