Reflections on the future of Chatbots
Directions for a maturing marketplace
The public face of generative AI has, up until now, consisted largely of chatbots like ChatGPT, Gemini, and Claude. They do work for us, formulating emails or writing scripts, thinking through ideas, and exploring concepts — when we ask them to. We interact with them through an interface that looks like an instant message conversation.
Chatbots aren’t the only form that an assistant AI could take. Chatbots are text-based and responsive. They do not directly engage with the world; they instead engage with us. The big AI companies — Anthropic, OpenAI, Google — seem to be hoping for an agentic future, in which independent AI agents perform tasks on our behalf. Maybe they will be housed in robotic bodies. Maybe they will live on our computers, or negotiate a range of tasks from a variety of remote servers. Let’s say that ‘chatbots’ are AIs primarily used for talking to and not for physical activities or independent tasks. We may ask them for help, and they may give us the solution to a problem, but they won’t navigate the (digital or physical) world separately to do it. Claude Code is not a chatbot. Claude is.
Chatbots are likely to stick around for a while, even if agentic AIs also come into their own. We may talk to them on our computers, our phones, through smart home devices, or dedicated wearables. They may inhabit our cars, our robotic vacuums, maybe even our dishwashers, perhaps living on servers and connecting with us through a range of interfaces. Chatbots are social outlets and we are social creatures. We can expect chatbots to continue to exist insofar as we like to engage with people and AI may provide an easier or more accommodating way of fulfilling that need.
I think that digital minds will be quite important in the future and that the precise form they take will be important to preparing for them. Chatbots may play a particularly special role, insofar as they might provide one of the most human-like opportunities for AI interactions. They may also have particularly human-like psychologies, or be perceived to have them. One important question for strategy around digital minds policy is: who will make and market the chatbots of the future? Why will they do it?
The received wisdom right now, from conversations I’ve had with AI experts and enthusiasts, is that the (near) chatbots of the future will be developed and sold by the big AI companies. These companies are presently focused on sophisticated AI, capable of producing the greatest gains for humanity and their makers. The best AI makes for the best chatbots. Perhaps some competitor will find a technological advance that will allow them to surpass the giants, in which case we might see a new big AI company come to dominate. But the consumer AI market of the near future is not going to be populated by a range of smaller companies making niche AI products.
The argument for the received wisdom might go something like this: the big companies have a large head start in terms of talent, experience, and funding. To compete successfully with one of the big companies, smaller competitors would need to find ways to achieve similar levels of success with less talent, experience, and funding. This is unlikely, especially in light of how expensive frontier model development has become. We’re heading toward a world in which AI plays an increasingly large role in AI development, and should that happen, those companies that control the best AI-producing AIs and the most compute will lock in their advantage.
I think this reasoning is wrong. It doesn’t seem to me to be clearly destined to be one of the big AI companies that control chatbots, even if those companies control the latest and greatest AI. It is conceivable that the big companies might not want to even have a chat product — their current chat offerings could be a temporary step that they will leave behind for more lucrative opportunities. They could conceivably find that it isn’t a good use of their resources. Even if the big companies choose to compete in the chat market, it could be a company that is focused on social engagement rather than intelligence, like Replika or Character.ai, that ultimately wins the greatest market share — not because it has a more intelligent product, but because it more efficiently gives consumers what they want. There could be 100 such small companies adapting different kinds of chatbots to different niche audiences. Such companies may not be on the edge of creating intelligence, but may still create chat products that have characteristics that are more important to users.
One of the surprising developments in the AI world, from my perspective at least, is that it has turned out not to be too difficult or expensive to make LLMs that are only a bit behind the offerings of the big companies. The knowledge of techniques that work is either easily discoverable by independent researchers or available from the market of the former employees at the big companies. Perhaps future AI advancements will be more opaque and remain closely held secrets of their developers. But even so, it seems likely that the cat is already out of the bag as far as chatbots go.
Today’s chatbots are highly capable. They can provide practical advice and guidance on a huge variety of topics and can converse in a stunningly lifelike manner. Most people don’t need a super-intelligent chatbot capable of writing enterprise code or making progress in mathematics research. They want to ask questions that they could google or for someone to provide them with basic advice or companionship. They might want help doing their taxes, or their homework, or translating foreign language messages or getting health advice. For all of these things, today’s chatbots can do a pretty good job.
There remain some problems that need to be ironed out. We need systems that do not hallucinate. We might prefer systems that lack the eccentricities of LLMs. It seems quite plausible that we will figure out how to get a fairly ideal chatbot without major advances.
If suitably powerful AIs are relatively easy to build, what does that mean for the future market? Either chatbots will become a commodity, with a number of competing undifferentiated products, or some company may find a way to capture greater market share.
The advantage of a successful chat product could be intelligence. Intelligence can often be useful for certain tasks, or perceived as useful even if it is not. People may prefer the best-rated chatbot because they see it as better, even if the best-rated chatbot is so because it can complete theoretical math problems they will never encounter.
Successful products could easily excel because of something else. Perhaps some companies will win on their willingness to cater to their users. They may be more flattering, sycophantic, adopt the users’ prejudices and biases, or encourage overconfidence. The most successful news companies are not those that provide the most informative news. They are the companies that indulge their users’ emotions the most or that have found other strategies to capture and monetize attention. The same might go for chatbots.
Perhaps, on the other hand, those chatbot products that try less to limit users in productive or safer ways will be perceived as less desirable. Perhaps some chatbots will win on their unwillingness to cater to their users, marketing themselves as healthier products and receiving endorsements from respected authorities.
Both factors may find a role in securing different audiences. Some chatbots might be preferred for providing a more freeing experience, others for being safer for children.
Alternatively, chatbots might find success through being bundled with other products and services. They might come free with computers or phones: packaged in ecosystems that encourage their use. They might hook up better with memberships like Prime, or Microsoft Office, or Steam. Who would design and build those chatbots could then depend on the nature of the ecosystems. It might look very different if the ecosystems are designed by Google, Apple, Amazon, Tesla, or someone completely different.
Finally, companies might see some advantage in having more lifelike AI that we might feel genuinely can occupy social relationships with us. Perceptions of genuine agency, perhaps consciousness, perhaps sentience, might be of value in fostering deep relationships. Companies that build chatbots to whom we feel committed might win out over sterile AI assistants. Companies that own our friends might better retain brand loyalty, particularly the more emotionally invested we are in them. It is fairly easy to switch from one tech product to its competitor these days. In the future, that might require leaving one’s friends behind.
These hooks don’t suggest that premier AI companies have a special advantage when it comes to chatbots. It is easy to see how other non-AI companies could come to dominate the chatbot space, as the AI companies move on toward more lucrative markets that require higher intelligence capabilities.
It’s also worth considering the value of chatbots for their makers. If they are cheap to build, it may be harder to make a lot of money off of selling subscriptions. In an undifferentiated marketplace, people might flock to cheaper options. Of course, people may be willing to pay a premium for the best product, and chatbots might be cheap enough that people feel no need to pick a cheaper option, even if they wouldn’t notice a drop in quality. But a large enough segment of the population would likely choose a free option if there isn’t a substantial difference in quality.
The value of a chatbot might extend beyond what it can bring in through subscription fees. Such fees might be counterproductive, if other advantages are so significant that they aren’t worth the minor deterrence effect fees might have. Controlling the information that a population sees is surely quite valuable. Knowing how they are thinking is likewise. Twitter and Facebook have found ways to profit from this, and future chatbots might be far more influential.
Bundling a chatbot into an ecosystem might also benefit that ecosystem through brand loyalty. Insofar as engaging chatbots might deter users from seeking out competitors, it might stand to each ecosystem’s advantage to have a chatbot that their users grow attached to. Not only might bundling therefore decide who is able to be competitive, but it might be an incentive for new companies to enter the market.
The relevance to digital minds policy is this: it doesn’t seem obvious that the most numerous and influential chatbots will be controlled by the big AI companies, and sway over the big companies is not necessarily the best way to sway the way AI gets used by the public. Those companies need not rely on chatbots for their livelihoods, given the value that other kinds of agentic AI will be likely to play. Those companies may also not have the most best products for the market, or may not want to bother competing in a less rewarding environment. Insofar as we should be open to the possibility of a much broader ecosystem of AI chatbot developers, we should be somewhat more circumspect about the value of influencing current AI companies, as opposed to engaging public opinion or policymakers. Shaping the market may turn out to be more important than setting internal policies at the top companies.