When Algorithms Think for Us: Who Shapes Humanity's Future, Us or Our Smart Models?

We write these lines as people who use AI every day. We build tools with it, review text with it, and test ideas with it. So we are not writing from fear, and not from awe. We are writing as people who see the tool up close and ask a simple question: as the tool gets smarter every year, who remains the one who decides?

This is not only a philosophical question. It is a practical one. The teacher grading homework written by a language model faces it. So does the doctor reading an automated recommendation, and the developer accepting a suggestion they didn't fully read. And we face it in the lab, too.

A map that changes every few months

Let's start with what we know, not what we imagine.

In a single year, model scores on hard benchmarks jumped sharply. Stanford's AI Index reports a gain of about 67 points on a benchmark of real-world software problems (SWE-bench) within a year of its release ^[2]. The same report notes that the cost of running a GPT-3.5-level system fell more than 280-fold between late 2022 and autumn 2024 ^[2]. What was expensive and rare became cheap and everywhere.

Adoption has kept pace. According to the same report, 78% of organizations said they used AI in 2024, up from 55% the year before ^[2]. These are no longer tools for big labs only. They sit in the office, the classroom and the phone.

And it's not just accuracy that is improving, but stamina. In a 2025 study, METR found that the length of tasks AI agents can complete with 50% success has roughly doubled every seven months over the previous six years ^[3]. METR itself notes that some of its figures age as new data comes in. But the direction is clear: the machine no longer just answers a question. It carries out a chain of steps.

Behind all of this is computing power growing fast. Epoch AI estimates that the training compute of notable models has doubled roughly every six months since 2010 ^[4].

What does this mean for us? It means today's map won't hold in a year. Anyone planning a school or an organization around what models can do right now will find themselves behind before they start.

When we hand over the compass without noticing

The nearest risk is not a machine rebelling against us. It is much simpler: that we stop thinking because something else is thinking for us.

In a study published in Radiology in 2023, 27 radiologists read mammograms alongside suggestions from an AI system. When the suggestion was correct, less experienced readers got it right about 80% of the time. When the suggestion was wrong, their accuracy dropped to 19.8% ^[5]. Even the most experienced readers fell from 82.3% to 45.5% ^[5]. Researchers call this "automation bias": we tend to believe the machine because it is a machine.

In a study by Microsoft Research presented at CHI 2025, 319 knowledge workers shared 936 examples from their daily work with generative AI ^[6]. The finding is striking: the more people trusted the tool, the less critical thinking they applied. The more they trusted their own ability in the task, the more critical thinking they applied ^[6].

In scouting, we know this lesson well. A compass doesn't walk for you. It tells you where north is, but you decide where to go, when to stop and when to turn back. A scout who follows the device without looking up at the trail can reach the edge of a valley feeling completely sure of himself.

Where is the line between empowerment and loss of control?

Empowerment is when the tool does what we couldn't, while we understand what it is doing and can stop it. Loss of control is when we can no longer understand, review, or say "no."

The problem is that the shift from one to the other doesn't happen in a single moment. It happens gradually, one small decision at a time.

The UN report Governing AI for Humanity acknowledges that no one today understands all of these systems' inner workings well enough to fully control their outputs or predict their evolution ^[8]. It also points to another gap: of seven prominent international AI governance initiatives outside the UN, only seven countries take part in all of them, while 118 countries take part in none, mostly in the Global South ^[8]. In other words, most of the world, our region included, is not at the table when the rules are drawn.

The International AI Safety Report, published in January 2025, brought together around 100 experts, with an advisory panel nominated by 30 countries plus the UN, the EU and the OECD ^[1]. The very existence of this effort says something: the question is no longer marginal.

On the legal side, the EU AI Act entered into force on 1 August 2024. It requires "appropriate human oversight" for high-risk systems in areas such as education and employment ^[9].

From our experience, we draw the line with these practical rules:

• If you can't explain the decision, don't rely on it. A tool that gives you an answer without a reason is asking for blind trust.
• A decision that affects a person passes through a person. Admission, evaluation, discipline and treatment are not left to the machine alone.
• Compute with logic, explain with AI. Whatever can be calculated with clear rules should not be left to a probabilistic model.
• Always keep the off switch. A system you can't stop or reverse is not a tool. It's a commitment.
• Watch your own skills. If you notice you can no longer do the task without the tool, that's a signal worth your attention.

What building our own tools has taught us

We didn't take these rules from books. We wrote them because we ran into them while building.

When we built Kashef, our security auditor, the easy option was to ask the model to give a website a score. We didn't. The score is computed with fixed logic that any person can check step by step. The AI's role is to explain the result in plain language, not to decide it. If the model gets the explanation wrong, the number is still correct and verifiable.

In LinkGuard, we decided not to stop at "this link is dangerous." The user always sees: why this verdict? Which signals did we find? That way people learn to see for themselves, and need the tool a little less next time.

And with Najem, the Telegram assistant serving more than 1,500 scout leaders, we learned that a good tool saves time but doesn't take the leader's place. It reminds, gathers and suggests. But judging an educational moment with a young scout at camp stays with a human who knows him.

What we can do now

There is a clear gap between experts and the public. Pew found in 2025 that 56% of AI experts expect AI to have a positive impact on the United States over the next 20 years, compared with only 17% of the public ^[7]. Yet both sides agree on one thing: around 55% of the public and 57% of experts want more control over how AI is used in their lives ^[7].

That shared wish is the starting point. And control isn't granted. It is practiced.

As individuals, we can read the answer before we copy it, and ask "how do you know?" before asking "what's the answer?". We can keep at least one skill we do with our own hands every week.

As educators and leaders, we can teach young people to use the tool and to challenge it at the same time. A simple activity: give the team an answer from a language model and ask them to find one mistake in it. This builds the muscle of healthy doubt. Another: ask them to solve the problem themselves first, then compare their answer with the machine's and discuss the difference. The tool becomes a colleague to review, not an authority to obey.

As builders, we can design systems that explain themselves, keep a human in the loop, and separate what is computed from what is suggested. Transparency isn't an extra feature. It's a condition for trust.

Our region doesn't have to be a consumer only. Every school, association or small team that builds a tool with clear values adds a voice to a table that is still being set.

Smart models will draw more accurate maps every year. That much is almost certain. But a map, however accurate, doesn't choose the destination.

The tool draws the map. We hold the compass.