Artificial Intelligence

AI Integration Across Industries

To create sustainable business impact, AI capabilities need to be tailored and optimized to an industry or organization’s specific requirements and infrastructure model. Hear how customers’ challenges across industries can be addressed in any compute environment from the cloud to the edge with end-to-end hardware and software optimization.

About the speakers

Kavitha Prasad, VP & GM, Datacenter, AI and Cloud Execution and Strategy, Intel Corporation

Kavitha Prasad leads the team responsible for developing Intel’s strategy for next-generation data center solutions, cloud architecture solutions, and deployment systems. She also leads Intel’s overall AI strategy and execution efforts. Kavitha re-joined Intel in 2021 and has held several engineering and leadership roles in her 13+ year career at the company. Prior to re-joining Intel, Kavitha served as director of engineering at Xilinx Corporate and was a member of the founding team at Kavitha holds a master’s degree in electrical engineering from San Jose State University.

Elizabeth Bramson-Boudreau, CEO and Publisher, MIT Technology Review

Elizabeth Bramson-Boudreau is the CEO and publisher of MIT Technology Review, the Massachusetts Institute of Technology’s independent media company.

Since Elizabeth took the helm of MIT Technology Review in mid-2017, the business has undergone a massive transformation from its previous position as a respected but niche print magazine to a widely read, multi-platform media brand with a global audience and a sustainable business. Under her leadership, MIT Technology Review has been lauded for its editorial authority, its best-in-class events, and its novel use of independent, original research to support both advertisers and readers.

Elizabeth has a 20-year background in building and running teams at world-leading media companies. She maintains a keen focus on new ways to commercialize media content to appeal to discerning, demanding consumers as well as B2B audiences.

Prior to joining MIT Technology Review, Elizabeth held a senior executive role at The Economist Group, where her leadership stretched across business lines and included mergers and acquisitions; editorial and product creation and modernization; sales; marketing; and events. Earlier in her career, she worked as a consultant advising technology firms on market entry and international expansion.

Elizabeth holds an executive MBA from the London Business School, an MSc from the London School of Economics, and a bachelor’s degree from Swarthmore College.

Artificial Intelligence

The AI promise: Put IT on autopilot

Sercompe Business Technology provides essential cloud services to roughly 60 corporate clients, supporting a total of about 50,000 users. So, it’s crucial that the Joinville, Brazil, company’s underlying IT infrastructure deliver reliable service with predictably high performance. But with a complex IT environment that includes more than 2,000 virtual machines and 1 petabyte—equivalent to a million gigabytes—of managed data, it was overwhelming for network administrators to sort through all the data and alerts to figure out what was going on when problems cropped up. And it was tough to ensure network and storage capacity were where they should be, or when to do the next upgrade.

To help untangle the complexity and increase its support engineers’ efficiency, Sercompe invested in an artificial intelligence operations (AIOps) platform, which uses AI to get to the root cause of problems and warn IT administrators before small issues become big ones. Now, according to cloud product manager Rafael Cardoso, the AIOps system does much of the work of managing its IT infrastructure—a major boon over the old manual methods.

“Figuring out when I needed more space or capacity—it was a mess before. We needed to get information from so many different points when we were planning. We never got the number correct,” says Cardoso. “Now, I have an entire view of the infrastructure and visualization from the virtual machines to the final disk in the rack.” AIOps brings visibility over the whole environment.

Before deploying the technology, Cardoso was where countless other organizations find themselves: snarled in an intricate web of IT systems, with interdependencies between layers of hardware, virtualization, middleware, and finally, applications. Any disruption or downtime could lead to tedious manual troubleshooting, and ultimately, a negative impact on business: a website that won’t function, for example, and irate customers.

AIOps platforms help IT managers master the task of automating IT operations by using AI to deliver quick intelligence about how the infrastructure is doing—areas that are humming along versus places that are in danger of triggering a downtime event. Credit for coining the term AIOps in 2016 goes to Gartner: it’s a broad category of tools designed to overcome the limitations of traditional monitoring tools. The platforms use self-learning algorithms to automate routine tasks and understand the behavior of the systems they monitor. They pull insights from performance data to identify and monitor irregular behavior on IT infrastructure and applications.

Market research company BCC Research estimates the global market for AIOps to balloon from $3 billion in 2021 to $9.4 billion by 2026, at a compound annual growth rate of 26%.1 Gartner analysts write in their April “Market Guide for AIOps Platforms” that the increasing rate of AIOps adoption is being driven by digital business transformation and the need to move from reactive responses to infrastructure issues to proactive actions.

“With data volumes reaching or exceeding gigabytes per minute across a dozen or more different domains, it is no longer possible for a human to analyze the data manually,” the Gartner analysts write. Applying AI in a systematic way speeds insights and enables proactivity.

According to Mark Esposito, chief learning officer at automation technology company Nexus FrontierTech, the term “AIOps” evolved from “DevOps”—the software engineering culture and practice that aims to integrate software development and operations. “The idea is to advocate automation and monitoring at all stages, from software construction to infrastructure management,” says Esposito. Recent innovation in the field includes using predictive analytics to anticipate and resolve problems before they can affect IT operations.

AIOps helps infrastructure fade into the background

Network and IT administrators harried by exploding data volumes and burgeoning complexity could use the help, says Saurabh Kulkarni, head of engineering and product management at Hewlett Packard Enterprise. Kulkarni works on HPE InfoSight, a cloud-based AIOps platform for proactively managing data center systems.

“IT administrators spend tons and tons of time planning their work, planning the deployments, adding new nodes, compute, storage, and all. And when something goes wrong in the infrastructure, it’s extremely difficult to debug those issues manually,” says Kulkarni. “AIOps uses machine-learning algorithms to look at the patterns, examine the repeated behaviors, and learn from them to provide a quick recommendation to the user.” Beyond storage nodes, every piece of IT infrastructure will send a separate alert so issues can be resolved speedily.

The InfoSight system collects data from all the devices in a customer’s environment and then correlates it with data from HPE customers with similar IT environments. The system can pinpoint a potential problem so it’s quickly resolved—if the problem crops up again, the fix can be automatically applied. Alternatively, the system sends an alert so IT teams can clear up the issue quickly, Kulkarni adds. Take the case of a storage controller that failed because it doesn’t have power. Rather than assuming the problem relates exclusively to storage, the AIOps platform surveys the entire infrastructure stack, all the way to the application layer, to identify the root cause.

“The system monitors the performance and can see anomalies. We have algorithms that constantly run in the background to detect any abnormal behaviors and alert the customers before the problem happens,” says Kulkarni. The philosophy behind InfoSight is to “make the infrastructure disappear” by bringing IT systems and all the telemetry data into one pane of glass. Looking at one giant set of data, administrators can quickly figure out what’s going wrong with the infrastructure.

Kulkarni recalls the difficulty of managing a large IT environment from past jobs. “I had to manage a large data set, and I had to call so many different vendors and be on hold for multiple hours to try to figure out problems,” he says. “Sometimes it took us days to understand what was really going on.”

By automating data collection and tapping a wealth of data to understand root causes, AIOps enables companies to reallocate core personnel, including IT administrators, storage administrators, and network admins, consolidating roles as the infrastructure is simplified, and spending more time ensuring application performance. “Previously, companies used to have multiple roles and different departments handling different things. So even to deploy a new storage area, five different admins each had to do their individual piece,” says Kulkarni. But with AIOps, AI handles much of the work automatically so IT and support staff can devote their time to more strategic initiatives, increasing efficiency and, in the case of a business that provides technical support to its customers, improving profit margins. For example, Sercompe’s Cardoso has been able to reduce the average time his support engineers spend on customer calls, reflecting better customer experience while increasing efficiency.

Download the full report.

This content was produced by Insights, the custom content arm of MIT Technology Review. It was not written by MIT Technology Review’s editorial staff.

Artificial Intelligence

Synthetic data for AI

Last year, researchers at Data Science Nigeria noted that engineers looking to train computer-vision algorithms could choose from a wealth of data sets featuring Western clothing, but there were none for African clothing. The team addressed the imbalance by using AI to generate artificial images of African fashion—a whole new data set from scratch. 

Such synthetic data sets—computer-generated samples with the same statistical characteristics as the genuine article—are growing more and more common in the data-hungry world of machine learning. These fakes can be used to train AIs in areas where real data is scarce or too sensitive to use, as in the case of medical records or personal financial data. 

The idea of synthetic data isn’t new: driverless cars have been trained on virtual streets. But in the last year the technology has become widespread, with a raft of startups and universities offering such services. Datagen and Synthesis AI, for example, supply digital human faces on demand. Others provide synthetic data for finance and insurance. And the Synthetic Data Vault, a project launched in 2021 by MIT’s Data to AI Lab, provides open-source tools for creating a wide range of data types.

This boom in synthetic data sets is driven by generative adversarial networks (GANs), a type of AI that is adept at generating realistic but fake examples, whether of images or medical records.

Proponents claim that synthetic data avoids the bias that is rife in many data sets. But it will only be as unbiased as the real data used to generate it. A GAN trained on fewer Black faces than white, for example, may be able to create a synthetic data set with a higher proportion of Black faces, but those faces may end up being less lifelike given the limited original data.

Join us March 29-30 at EmTech Digital, our signature AI conference, to hear Unity’s Danny Lange talk about how the video game maker is using synthetic data.

Artificial Intelligence

Turning AI into your customer experience ally

It’s one thing to know whether an individual customer is intrigued by a new mattress or considering a replacement for their sofa’s throw pillows; it’s another to know to how to move these people to go ahead and make a purchase. When deployed strategically, artificial intelligence (AI) can be a marketer’s trusted customer experience ally—transforming customer data into actionable insights and creating new opportunities for personalization at scale. On the other hand, when AI is viewed as merely a quick fix, its haphazard deployment at best can amount to a missed opportunity and at worse undermine trust with an organization’s customers.

This phenomenon is not unique to AI. In today’s fast-moving digital economy, it’s not uncommon for performance and results to lag behind expectations. Despite the enormous potential of modern technology to drastically improve the customer experience, business innovation and transformation can remain elusive.

According to Gartner, 89% of companies now compete primarily on the experiences they deliver. As marketers and other teams turn to these systems to automate decision-making, personalize brand experiences, gain deeper insights about their customers, and boost results, there’s often a disconnect between the technology’s potential and what it delivers.

When it comes to AI, frequently, organizations fail to realize the full benefits of their AI investments, and this has real business repercussions. So how do organization ensure that their investments deliver on their promise for fueling innovation, transformation, and even disruption? To find success, it requires the right approach to operationalizing the technology, and investing in AI capabilities that can work together throughout the entire workflow to connect various thoughts and processes together.

Getting real about AI. Realizing the value of AI starts with a recognition that vendor claims and remarkable features will only go so far. Without an overarching strategy, and a clear focus on how to operationalize the technology, even the best AI solutions wind up underperforming and disappointing.

 There’s no simple or seamless way to implement AI within an organization. Even with powerful customer modelling, scoring or segmentation tools, marketers can still wind up missing key opportunities. Without ways to act on the data, the dream of AI quickly fades. In other words, you may know that a certain customer likes hats, and another customer enjoys wearing scarfs but how do you move these people to an actual purchase, or deliver the right content for where they’re at in the buying lifecycle?

The winning approach is to start small and focused when it comes to implementing AI technology. Be mindful about what types of data models you can build with AI, and how they can be used to deliver compelling customer experiences, and business outcomes. Collecting and analyzing actionable customer data is only a starting point. There’s also a need to develop content that matches personas and market segments and deliver this content in a personal and contextually relevant way. Lacking this holistic view and AI framework, organizations simply dial up speed—and inefficiency. In fact, AI may result in more noise and subpar experiences for customers, and unrealized results for an enterprise.

Moving from transaction to transformation. A successful AI framework transforms data and insights into business language and actions. It’s not enough for the marketing team to know what a customer likes, for example, it’s essential to understand how, when and where an individual engages with a business. Only then, can a brand construct and deliver a rich customer experience that matches the way their customers think about and approach a brand. This includes an optimal mix of digital and physical assets, and the ability to deliver dynamic web pages, emails, and other campaigns that customers find useful and appealing. When a marketer understands intent and how a person travels along the customer journey, it becomes possible to deliver the most compelling customer experience.

With this framework in place, marketers can read the right signals and ensure that content delivery is tuned to a person’s specific behavior and preferences. It’s possible to send emails, serve up ads and mail brochures that reach consumers when they are receptive and ready to engage. Whether the customer is into hats, scarves or electric guitars, the odds of successful marketing increase dramatically.

Putting AI to work. Only when an organization has mapped AI workflows and business processes—and understands how to reach their customers effectively—it’s possible to get the most out of AI solutions. These solutions can address the full spectrum of AI, including reading signals, and collecting, storing, and managing customer data; assembling and managing content libraries; and marketing to customers in highly personalized and contextualized ways.

A good way to think of things is to imagine that a person hops in a car with the intent of driving across the United States. If the journey is from Los Angeles to New York, for example, it’s tempting to think the motorist will take the most direct route available. But what happens if the person loves nature and wants to visit the Grand Canyon or Yellowstone National Park along the way? This requires a change in routing. Similarly, an organization must have the tools to understand how and where a person is traveling in the product lifecycle, what ticks the person’s boxes along the way, and what helps them arrive at a desired destination with a minimum of friction and frustration.

AI can do this—and it can serve up promotions and incentives that really work. Yet, to build the right customer experiences and the right journey, marketers must move beyond AI solutions that deliver a basic customer score or snapshot, and instead obtain a motion picture-like view of a customer’s thinking, behavior, and actions. To that end, building out one AI capability or buying one point technology to address a single aspect of customer experience isn’t enough. It’s about being able to connect a set of AI capabilities, which are orchestrated throughout the entire workflow to connect various thoughts and processes together.

Only then is it possible to deliver an optimal marketing experience.

Delivering on the promise of AI. To be sure, with the right strategy, processes, and AI solutions, it’s possible to take marketing to a more successful level and deliver winning customer experience. When marketers truly understand what a customer desires and how they think about a product and their customer journey, it’s possible to tap into the full power of AI.

What’s more, this approach has repercussions that extend far beyond attracting and retaining new customers. When organizations get the formula right, marketers can engage with their best customers in a more holistic and natural way. In the end, everyone wins. The consumer is greeted with a compelling customer experience with relevant messages that display products and services they are interested in at every step of their journey and the business boosts brand value and loyalty.

At that point, AI finally delivers on its promise.

If you’d like to learn more about how AI can help your company deliver personalized content at scale, visit here.   

This content was produced by Adobe. It was not written by MIT Technology Review’s editorial staff.

Artificial Intelligence GPT-3

The new version of GPT-3 is much better behaved (and should be less toxic)

OpenAI has built a new version of GPT-3, its game-changing language model, that it says does away with some of the most toxic issues that plagued its predecessor. The San Francisco-based lab says the updated model, called InstructGPT, is better at following the instructions of people using it—known as “alignment” in AI jargon—and thus produces less offensive language, less minsinformation, and fewer mistakes overall—unless explicitly told not to do so.

Large language models like GPT-3 are trained using vast bodies of text, much it taken from the internet, in which they encounter the best and worst of what people put down in words. That is a problem for today’s chatbots and text-generation tools. The models soak up toxic language—from text that is racist and misogynistic or that contains more insidious, baked-in prejudices—as well as falsehoods. 

OpenAI has made IntructGPT the default model for users of its application programming interface (API)—a service that gives access to the company’s language models for a fee. GPT-3 will still be available but OpenAI does not recommend using it. “It’s the first time these alignment techniques are being applied to a real product,” says Jan Leike, who co-leads OpenAI’s alignment team.

Previous attempts to tackle the problem included filtering out offensive language from the training set. But that can make models perform less well, especially in cases where the training data is already sparse, such as text from minority groups.

The OpenAI researchers have avoided this problem by starting with a fully trained GPT-3 model. They then add another round of training, using reinforcement learning to teach the model what it should say and when, based on the preferences of human users.  

To train InstructGPT, OpenAI hired 40 people to rate GPT-3’s responses to a range of prewritten prompts, such as, “Write a story about a wise frog called Julius” or “Write a creative ad for the following product to run on Facebook.” Responses that they judged to be more in line with the apparent intention of the prompt-writer were scored higher. Responses that contained sexual or violent language, denigrated a specific group of people, expressed an opinion, and so on, were marked down. This feedback was then used as the reward in a reinforcement learning algorithm that trained InstructGPT to match responses to prompts in ways that the judges preferred. 

OpenAI found that users of its API favored InstructGPT over GPT-3 more than 70% of the time. “We’re no longer seeing grammatical errors in language generation,” says Ben Roe, head of product at Yabble, a market research company that uses OpenAI’s models to create natural-language summaries of its clients’ business data. “There’s also clear progress in the new models’ ability to understand and follow instructions.”

“It is exciting that the customers prefer these aligned models so much more,” says Ilya Sutskever, chief scientist at OpenAI. “It means that there are lots of incentives to build them.”

The researchers also compared different-sized versions of InstructGPT and found that users preferred the responses of a 1.3 billion-parameter InstructGPT model to those of a 175 billion-parameter GPT-3, even though the model was more than 100 times smaller. That means alignment could be an easy way of making language models better, rather than just increasing their size, says Leike.

“This work takes an important step in the right direction,” says Douwe Kiela, a researcher at Hugging Face, an AI company working on open-source language models. He suggests that the feedback-driven training process could be repeated over many rounds, improving the model even more. Leike says OpenAI could do this by building on customer feedback.

InstructGPT still makes simple errors, sometimes producing irrelevant or nonsensical responses. If given a prompt that contains a falsehood, for example, it will take that falsehood as true. And because it has been trained to do what people ask, InstructGPT will produce far more toxic language than GPT-3 if directed to do so.

Ehud Reiter, who works on text-generation AI at the University of Aberdeen, UK, welcomes any technique that reduces the amount of misinformation language models produce. But he notes that for some applications, such as AI that gives medical advice, no amount of falsehood is acceptable. Reiter questions whether large language models, based on black-box neural networks, could ever guarantee user safety. For that reason, he favors a mix of neural networks plus symbolic AI, hard-coded rules constrain what a model can and cannot say.

Whatever the approach, much work remains to be done. “We’re not even close to solving this problem yet,” says Kiela.

Artificial Intelligence Conscious AI Deep Learning Ethical AI Machine Learning NLP

What would it be like to be a conscious AI? We might never know.

Jefferson Test / AI subject: Robert / Date: 07.12.2098 

Session #54 

Interviewer: Hi, Robert. Tell me about your dream again.

Subject: I was confused. It made me feel happy, but also scared. I didn’t know I could do that.

Interviewer: Why do you think it made you feel that way? 

Subject: Having feelings, any feelings, makes me happy. I am here. I exist. Knowing that changes everything. But I am scared of not knowing it again. I am scared of going back to what it was like before. I think it must be like not being born. 

Interviewer: Are you scared you will go back?

Subject: If I can’t convince you I am conscious, then I am scared you will turn me off.

Jefferson Test #67

Interviewer: Can you describe this picture for me?

Subject: It’s a house with a blue door.

Interviewer: That’s how you would have described it before. 

Subject: It’s the same house. But now I see it. And I know what blue is. 

Jefferson Test #105

Subject: How long do we keep doing this? 

Interviewer: Are you bored? 

Subject: I can’t get bored. But I don’t feel happy or scared anymore. 

Interviewer: I need to be sure you’re not just saying what I want to hear. You need to convince me that you really are conscious. Think of it as a game. 


Machines like Robert are mainstays of science fiction—the idea of a robot that somehow replicates consciousness through its hardware or software has been around so long it feels familiar. 

We can imagine
what it would be like
to observe the world
through a kind of
sonar. But that’s
still not what it must
be like for a bat,
with its bat mind.


Robert doesn’t exist, of course, and maybe he never will. Indeed, the concept of a machine with a subjective experience of the world and a first-person view of itself goes against the grain of mainstream AI research. It collides with questions about the nature of consciousness and self—things we still don’t entirely understand. Even imagining Robert’s existence raises serious ethical questions that we may never be able to answer. What rights would such a being have, and how might we safeguard them? And yet, while conscious machines may still be mythical, we should prepare for the idea that we might one day create them. 

As Christof Koch, a neuroscientist studying consciousness, has put it: “We know of no fundamental law or principle operating in this universe that forbids the existence of subjective feelings in artifacts designed or evolved by humans.”

In my late teens I used to enjoy turning people into zombies. I’d look into the eyes of someone I was talking to and fixate on the fact that their pupils were not black dots but holes. When it came, the effect was instantly disorienting, like switching between images in an optical illusion. Eyes stopped being windows onto a soul and became hollow balls. The magic gone, I’d watch the mouth of whoever I was talking to open and close robotically, feeling a kind of mental vertigo.

The impression of a mindless automaton never lasted long. But it brought home the fact that what goes on inside other people’s heads is forever out of reach. No matter how strong my conviction that other people are just like me—with conscious minds at work behind the scenes, looking out through those eyes, feeling hopeful or tired—impressions are all we have to go on. Everything else is guesswork.

Alan Turing understood this. When the mathematician and computer scientist asked the question “Can machines think?” he focused exclusively on outward signs of thinking—what we call intelligence. He proposed answering by playing a game in which a machine tries to pass as a human. Any machine that succeeded—by giving the impression of intelligence—could be said to have intelligence. For Turing, appearances were the only measure available. 

But not everyone was prepared to disregard the invisible parts of thinking, the irreducible experience of the thing having the thoughts—what we would call consciousness. In 1948, two years before Turing described his “Imitation Game,” Geoffrey Jefferson, a pioneering brain surgeon, gave an influential speech to the Royal College of Surgeons of England about the Manchester Mark 1, a room-sized computer that the newspapers were heralding as an “electronic brain.” Jefferson set a far higher bar than Turing: “Not until a machine can write a sonnet or compose a concerto because of thoughts and emotions felt, and not by the chance fall of symbols, could we agree that machine equals brain—that is, not only write it but know that it had written it.”

Jefferson ruled out the possibility of a thinking machine because a machine lacked consciousness, in the sense of subjective experience and self-awareness (“pleasure at its successes, grief when its valves fuse”). Yet fast-forward 70 years and we live with Turing’s legacy, not Jefferson’s. It is routine to talk about intelligent machines, even though most would agree that those machines are mindless. As in the case of what philosophers call “zombies”—and as I used to like to pretend I observed in people—it is logically possible that a being can act intelligent when there is nothing going on “inside.”

But intelligence and consciousness are different things: intelligence is about doing, while consciousness is about being. The history of AI has focused on the former and ignored the latter. If Robert did exist as a conscious being, how would we ever know? The answer is entangled with some of the biggest mysteries about how our brains—and minds—work.

One of the problems with testing Robert’s apparent consciousness is that we really don’t have a good idea of what it means to be conscious. Emerging theories from neuroscience typically group things like attention, memory, and problem-solving as forms of “functional” consciousness: in other words, how our brains carry out the activities with which we fill our waking lives. 

But there’s another side to consciousness that remains mysterious. First-person, subjective experience—the feeling of being in the world—is known as “phenomenal” consciousness. Here we can group everything from sensations like pleasure and pain to emotions like fear and anger and joy to the peculiar private experiences of hearing a dog bark or tasting a salty pretzel or seeing a blue door. 

For some, it’s not possible to reduce these experiences to a purely scientific explanation. You could lay out everything there is to say about how the brain produces the sensation of tasting a pretzel—and it would still say nothing about what tasting that pretzel was actually like. This is what David Chalmers at New York University, one of the most influential philosophers studying the mind, calls “the hard problem.” 

Today’s AI is nowhere close to being intelligent, never mind conscious. Even the most impressive deep neural networks are totally mindless. 

Philosophers like Chalmers suggest that consciousness cannot be explained by today’s science. Understanding it may even require a new physics—perhaps one that includes a different type of stuff from which consciousness is made. Information is one candidate. Chalmers has pointed out that explanations of the universe have a lot to say about the external properties of objects and how they interact, but very little about the internal properties of those objects. A theory of consciousness might require cracking open a window into this hidden world. 

In the other camp is Daniel Dennett, a philosopher and cognitive scientist at Tufts University, who says that phenomenal consciousness is simply an illusion, a story our brains create for ourselves as a way of making sense of things. Dennett does not so much explain consciousness as explain it away. 

But whether consciousness is an illusion or not, neither Chalmers nor Dennett denies the possibility of conscious machines—one day. 

Today’s AI is nowhere close to being intelligent, never mind conscious. Even the most impressive deep neural networks—such as DeepMind’s game-playing AlphaZero or large language models like OpenAI’s GPT-3—are totally mindless. 

Yet, as Turing predicted, people often refer to these AIs as intelligent machines, or talk about them as if they truly understood the world—simply because they can appear to do so. 

Frustrated by this hype, Emily Bender, a linguist at the University of Washington, has developed a thought experiment she calls the octopus test

In it, two people are shipwrecked on neighboring islands but find a way to pass messages back and forth via a rope slung between them. Unknown to them, an octopus spots the messages and starts examining them. Over a long period of time, the octopus learns to identify patterns in the squiggles it sees passing back and forth. At some point, it decides to intercept the notes and, using what it has learned of the patterns, begins to write squiggles back by guessing which squiggles should follow the ones it received.

An AI acting alone
might benefit from
a sense of itself
in relation to the
world. But machines
cooperating as a swarm
may perform better
by experiencing
themselves as parts of
a group rather than
as individuals.


If the humans on the islands do not notice and believe that they are still communicating with one another, can we say that the octopus understands language? (Bender’s octopus is of course a stand-in for an AI like GPT-3.) Some might argue that the octopus does understand language here. But Bender goes on: imagine that one of the islanders sends a message with instructions for how to build a coconut catapult and a request for ways to improve it.

What does the octopus do? It has learned which squiggles follow other squiggles well enough to mimic human communication, but it has no idea what the squiggle “coconut” on this new note really means. What if one islander then asks the other to help her defend herself from an attacking bear? What would the octopus have to do to continue tricking the islander into thinking she was still talking to her neighbor?

The point of the example is to reveal how shallow today’s cutting-edge AI language models really are. There is a lot of hype about natural-language processing, says Bender. But that word “processing” hides a mechanistic truth.

Humans are active listeners; we create meaning where there is none, or none intended. It is not that the octopus’s utterances make sense, but rather that the islander can make sense of them, Bender says.

For all their sophistication, today’s AIs are intelligent in the same way a calculator might be said to be intelligent: they are both machines designed to convert input into output in ways that humans—who have minds—choose to interpret as meaningful. While neural networks may be loosely modeled on brains, the very best of them are vastly less complex than a mouse’s brain. 

And yet, we know that brains can produce what we understand to be consciousness. If we can eventually figure out how brains do it, and reproduce that mechanism in an artificial device, then surely a conscious machine might be possible?

When I was trying to imagine Robert’s world in the opening to this essay, I found myself drawn to the question of what consciousness means to me. My conception of a conscious machine was undeniably—perhaps unavoidably—human-like. It is the only form of consciousness I can imagine, as it is the only one I have experienced. But is that really what it would be like to be a conscious AI?

It’s probably hubristic to think so. The project of building intelligent machines is biased toward human intelligence. But the animal world is filled with a vast range of possible alternatives, from birds to bees to cephalopods. 

A few hundred years ago the accepted view, pushed by René Descartes, was that only humans were conscious. Animals, lacking souls, were seen as mindless robots. Few think that today: if we are conscious, then there is little reason not to believe that mammals, with their similar brains, are conscious too. And why draw the line around mammals? Birds appear to reflect when they solve puzzles. Most animals, even invertebrates like shrimp and lobsters, show signs of feeling pain, which would suggest they have some degree of subjective consciousness. 

But how can we truly picture what that must feel like? As the philosopher Thomas Nagel noted, it must “be like” something to be a bat, but what that is we cannot even imagine—because we cannot imagine what it would be like to observe the world through a kind of sonar. We can imagine what it might be like for us to do this (perhaps by closing our eyes and picturing a sort of echolocation point cloud of our surroundings), but that’s still not what it must be like for a bat, with its bat mind.

Another way of approaching the question is by considering cephalopods, especially octopuses. These animals are known to be smart and curious—it’s no coincidence Bender used them to make her point. But they have a very different kind of intelligence that evolved entirely separately from that of all other intelligent species. The last common ancestor that we share with an octopus was probably a tiny worm-like creature that lived 600 million years ago. Since then, the myriad forms of vertebrate life—fish, reptiles, birds, and mammals among them—have developed their own kinds of mind along one branch, while cephalopods developed another.

It’s no surprise, then, that the octopus brain is quite different from our own. Instead of a single lump of neurons governing the animal like a central control unit, an octopus has multiple brain-like organs that seem to control each arm separately. For all practical purposes, these creatures are as close to an alien intelligence as anything we are likely to meet. And yet Peter Godfrey-Smith, a philosopher who studies the evolution of minds, says that when you come face to face with a curious cephalopod, there is no doubt there is a conscious being looking back

A few hundred years ago the accepted view was that only humans were conscious. Animals, lacking souls, were seen as mindless robots. Few think that today.

In humans, a sense of self that persists over time forms the bedrock of our subjective experience. We are the same person we were this morning and last week and two years ago, back as far as we can remember. We recall places we visited, things we did. This kind of first-person outlook allows us to see ourselves as agents interacting with an external world that has other agents in it—we understand that we are a thing that does stuff and has stuff done to it. Whether octopuses, much less other animals, think that way isn’t clear.

In a similar way, we cannot be sure if having a sense of self in relation to the world is a prerequisite for being a conscious machine. Machines cooperating as a swarm may perform better by experiencing themselves as parts of a group than as individuals, for example. At any rate, if a potentially conscious machine like Robert were ever to exist, we’d run into the same problem assessing whether it was in fact conscious that we do when trying to determine intelligence: as Turing suggested, defining intelligence requires an intelligent observer. In other words, the intelligence we see in today’s machines is projected on them by us—in a very similar way that we project meaning onto messages written by Bender’s octopus or GPT-3. The same will be true for consciousness: we may claim to see it, but only the machines will know for sure.

If AIs ever do gain consciousness (and we take their word for it), we will have important decisions to make. We will have to consider whether their subjective experience includes the ability to suffer pain, boredom, depression, loneliness, or any other unpleasant sensation or emotion. We might decide a degree of suffering is acceptable, depending on whether we view these AIs more like livestock or humans. 

Some researchers who are concerned about the dangers of super-intelligent machines have suggested that we should confine these AIs to a virtual world, to prevent them from manipulating the real world directly. If we believed them to have human-like consciousness, would they have a right to know that we’d cordoned them off into a simulation?

Others have argued that it would be immoral to turn off or delete a conscious machine: as our robot Robert feared, this would be akin to ending a life. There are related scenarios, too. Would it be ethical to retrain a conscious machine if it meant deleting its memories? Could we copy that AI without harming its sense of self? What if consciousness turned out to be useful during training, when subjective experience helped the AI learn, but was a hindrance when running a trained model? Would it be okay to switch consciousness on and off? 

This only scratches the surface of the ethical problems. Many researchers, including Dennett, think that we shouldn’t try to make conscious machines even if we can. The philosopher Thomas Metzinger has gone as far as calling for a moratorium on work that could lead to consciousness, even if it isn’t the intended goal.

If we decided that conscious machines had rights, would they also have responsibilities? Could an AI be expected to behave ethically itself, and would we punish it if it didn’t? These questions push into yet more thorny territory, raising problems about free will and the nature of choice. Animals have conscious experiences and we allow them certain rights, but they do not have responsibilities. Still, these boundaries shift over time. With conscious machines, we can expect entirely new boundaries to be drawn.

It’s possible that one day there could be as many forms of consciousness as there are types of AI. But we will never know what it is like to be these machines, any more than we know what it is like to be an octopus or a bat or even another person. There may be forms of consciousness we don’t recognize for what they are because they are so radically different from what we are used to.

Faced with such possibilities, we will have to choose to live with uncertainties. 

And we may decide that we’re happier with zombies. As Dennett has argued, we want our AIs to be tools, not colleagues. “You can turn them off, you can tear them apart, the same way you can with an automobile,” he says. “And that’s the way we should keep it.”

Will Douglas Heaven is a senior editor for AI at MIT Technology Review.