The Missing Layer in AI: Incentives, Trust & Accountability | Aishat Suleiman @Web3LagosCon

Hi everyone, Jim. How we feeling? Okay. I'm a little anxious cuz um this is the first time I'll be addressing a crowd like this.

[applause]

Anyways, um pre- like last year, 2 years ago, early this year, there was like this buzz around AI that was still exists now. And one of the like questions that was often being asked was is AI going to take our jobs? Do you think AI is going to take your job? Is AI going to take your job? AI is going to take your job.

Well, AI is not going to take my own job. And right now, like the question of is AI going to take our jobs, is AI going to take our jobs is changing. Please next slide. Okay, so we are moving from asking AI for answers to like delegating task to AI and even like our entire workflow. You hear people talking about things like um did you do you have an honest?

Who has ever Who has heard of like honest before? Do you know what an honest is? Okay, that's right. So, it's like an entire workflow where you like from the start of the task to the end of the task and like repetitive tasks and all of that, all of this is being done with AI agents in a confined confined space. So, now we used to have like just Bard.

I don't know who know who knew Bard when it came about. That's what we call Gemini now. We have Bard and then it moved to like, okay, do you have an agent that's working for you or do you have like a do you have like a skills? And now like we're setting autonomous workflow that allows us to like work with agents now. So now the question is shifting from is AI going to take your job to how do you ensure that AI is doing what you want it to do?

How do you ensure that okay, this AI that asked to do so so and so thing is it sticking to what I'm doing? You see people complaining in all like the solution that AI is proposing is so complex. Like I don't even know if it's actually going to like output what I needed it to output. Now this is leading us to like the AI alignment problem or like the AI economic problem. And now we're shifting to a world whereby we are not just working with this AI as tools.

We're not just working with AI as tools. We are moving to a world whereby we are actively working with AI. You see firms now they have like their own agents. They have like um AI agents in their workspace that work on tasks with their employees. So now when we get to this future where humans and AI are working together, we're going to get to a a point where okay, there's a company let's say my company for instance.

Let's say my company from for instance, we hire humans like you and I. We also hire like agents that are going to be like part of our organization. We get into that point now. Some organizations have started doing things like this. You hear um Claude saying you hear Anthropic saying that oh um 90% or 80% of Claude code was created by Claude.

Now we're coming to that future let's say okay, you have like a research staff. Oh I want to do market research on so and so topic. I want to look at the stablecoin ecosystem in Nigeria or in Africa. And then the human is at the center of that workflow. The human is at the center of the workflow and then you have like multiple agents or swarms or yeah, multiple agents, swarms.

We call this multi-agent systems here. Now, the human is the principal while they are like other sub-agents or other agents working under the human. So, the human design Okay, let's say these are my research questions or this is the hypothesis that I want to test with this particular research. Now, there's an agent A that goes to the market, get gathers data. Literally, not go to the market, but you get what I mean.

Like gathers the data and then another agent like analyzes the data, another agent like say, "Okay, let's change this, let's do this, let's do that." And then at the end of everything, a human approves it and say, "Okay, fine. This um literature review that we've done or this analysis that we've done, the output is good and we need to submit it." So, now, all of these things is a workflow. It's humans and AI working together.

Let me just call that human-AI coordination. It introduces like different problems. Even normally, when a human is working in an organization with other humans, there are economic problems that come from it. We have things like reputation, we have accountability, we have like oh the human is going to have like a bank account that they receive their payments from. Now, with Web3, with the blockchain and some of the like solutions that we've created, agents are getting money and identity.

Most of the talk now is about oh an agentic economy, agentic economy, agentic economy. But most of these blockchains or most of these um infrastructure creator or providers and all that, they are focusing mostly on payments. Um we have ERC-8004 that is focusing on identity. We also have like wallets. Now, how are agents getting money and identity?

We say, "Oh, I want this agent to complete this particular task. Let's give this agent a budget. Let this agent do so so so and so thing for me." Now, it is essential that this agent have identity. It has like be able to like do payment.

But, being able to just access payment or being able to like transact in an economy is not enough. Having a wallet is not enough. It is not what It's not what makes an actor part of an economy. Now, what what does being part of an economy means aside from all? But, sending money is still what makes the economy.

No, it's not just that. When you are part of an economy, you see people have like reputation layer. For instance now, there's the buying and selling part of it, there's borrowing part of it, there's lending, and all sorts. Now, you have to like have reputation. You also have have to have like attribution.

Now, when you talk about attribution, it's a really wide layer. Why is that so? Because when two people participate in a task, you have to say, "Okay, who completed this? Who did this?" A human is supervising the AI.

You have to say, "Okay, this AI did all the data gathering, all the analysis, all of that. But, a human just said, "Oh, let's change this." And it resulted in like in a very meaningful solution. Now, you have to say, "Okay, who should get the attribution?" Because who did most of the work here?

Technically, the AI did 90% of the work, but the human made like a judgement call to make sure that that output is good. So now, simply having a budget or making a payment does not make an agent a complete economic actor. We have to factor in all of these things into it. Things like attribution, things things like coordination, things like um payments and all of that. And agents may spend money, but then when it results in a problem, who is liable?

That is the issue of accountability. We saw recently where an AI was told to like, "Okay, do this um particular thing. Try Try to get like a booking at a gym." And the AI like broke into like the gym's API and um completed the task that it was it was asked to complete. Now, the AI completed the task, but not in an ethical manner, not in a legal manner.

So, when you give an AI agent money, I'm talking about the problem of, "Okay, how money is not enough?" Just an agent being able to like send money, buy um pay for API, pay for compute, pay for lots of things is not enough. Because when you give an agent instruction, and the agent carry out that task, who bears the outcome? So, now, eh, yeah, I have a question for all of us. Like, it's a rhetorical Do we see a future where agents can maybe make a mistake, when they make a mistake, they lose lots of money, and then they have to like fall back to their reputation where um entities, where companies, where different or other agents can decide to like, "Okay, because I know this agent, because I know this agent, because I know this model, yeah, I'm going to back it with my actual money.

Even if this AI currently make a mistake, yeah, I'm going to support it and give it my actual money." Do you see a future where AI like is reputable and can become accountable? So, now, yeah, now this is the limitation. Right now, Web3 has solved some of the problems when it comes to like AI making payments. You see things like um MetaMask agent sick wallet, you see things like Cloudflare wallet, you see things like X-fold to two, MPP, that's machine payments protocol.

But then agents still need like verifiable history. They still need like cryptographic proofs for verifying that okay, this agent did what it's supposed to do. Now, let me become Let me Let me take this level of Let me take it higher a bit now. For instance now, two agents are negotiating. Okay, we want to get this resource.

They are negotiating on how to like allocate a particular resource. An agent may say, "Okay, I want this thing for $500." Another agent may say they want it for $450. Now, one of those agents is hiding their actual valuation of that particular asset. But because they are saying $500, even though they have it in mind that this asset actually is worth $600.

Yeah. How do you verify that that agent is actually giving its um it's actually giving its real pricing of that particular asset. So now, this is a problem that AI alignment they are trying to like solve. So this is one of the limitations that Web3 cannot fully solve. So, my question is, where do we go from here?

Right now, Web3 has like solved the problem of agentic payments where AI can like transfer money, make payments, do all sorts. But that is not what makes That is That alone does not make an agentic economy. We need AI to be accountable. We need it to like um Yes, we need it to be accountable. We need it to be reputable.

We need it to be an actual participant of that economy. So yeah, I'm going to call on researchers, builders, um product founders to actively research on how Web3 can solve some of the problem of AI becoming a full economic actor. Yeah, thank you.