The seven deadly sins of product design, and what they miss

Reid Hoffman says social networks do best when they tap into one of the seven deadly sins. The sins explain why someone signs up. They explain much less about what happens to that person afterward, and that gap widens once AI agents start acting for their users.

In 2011, Reid Hoffman gave the Wall Street Journal a rule of thumb for the consumer internet. “Social networks do best when they tap into one of the Seven Deadly Sins,” said Hoffman, who co-founded LinkedIn and was an early investor in Facebook. “Facebook is [pride]. Zynga is sloth. LinkedIn is greed.” The bracket comes from his own retelling in a 2021 essay, ten years later.

Hoffman began angel investing in the dot-com era, and the sins became “one criterion for my investing decisions.” At Y Combinator’s Startup School in 2016, he told Sam Altman why. Most people he talked to, he said, “can’t articulate a robust theory of human nature.” The sins pushed founders toward one question. “If you’re trying to create a mass market consumer application, what parts of what it is to be human are you actually triggering?”

The question catches a habit common among engineers and researchers, who tend to judge a product by what its technology can do. Users judge it by whether it gets them something they already want. The sins answer half the question. The other half is what a product does to people once it has them, and Hoffman’s own account of building LinkedIn is mostly about that half.

A rogues’ gallery

Each sin names a desire much older than the internet. On Hoffman’s account, a product succeeds when it finds a new way to satisfy an old one.

Pride

“Look at me.”

Facebook Hoffman

Greed

“What’s in it for me?”

LinkedIn Hoffman

Lust

“Swipe right.”

Tinder and Hinge

Envy

“Why them?”

Instagram

Gluttony

“Just one more.”

Autoplay and endless feeds

Wrath

“Someone is wrong on the internet!”

Twitter Hoffman

Sloth

“Do it for me.”

Zynga Hoffman

“What parts of what it is to be human are you actually triggering?”

Reid Hoffman, Startup School, 2016

Figure 1. The seven deadly sins as desires a product can feed. Cards marked Hoffman carry his own pairings from 2011 and 2021. The other examples are illustrative.

Several of the sins are milder than their names. Hoffman’s greed is the pursuit of opportunity, and financial security pays for plenty of good things. Envy can spoil an evening spent on other people’s holiday photos. It can also show someone from a similar background that a goal is within reach.

Sloth, in product terms, is a preference for less effort, which is often the rational choice. Uber and food delivery are built on it. It is also the sin that AI agents will feed most directly.

Wrath is on the list because Hoffman watched one platform change sins. He first thought people used Twitter out of pride, to let everyone know they were having dinner at the hot new restaurant. “But, over time, venom and outrage overtook humblebrags,” he wrote. “So, wrath started to dominate as the primary sin there.”

Greed wrote 90 characters

Hoffman’s 2021 essay, Investing in Vices and Virtues, first ran on the Knight Foundation’s site. It records what happened when the heuristic met LinkedIn’s first users. Before launch, Hoffman expected greed to produce detailed profiles. A complete profile makes a member easier to find, and being found is how a professional network pays off.

Early members wrote almost nothing. “At one point in our first few years, the average profile length was just 90 characters,” Hoffman wrote. That was “shorter than a tweet!”

The sin was real. Members did want better careers. The wanting stopped short of typing.

Pride, a sin nobody had designed for, filled the gap. LinkedIn showed each member’s connection count, and early users were soon “amassing 1,000, 5,000 and even 20,000 connections.” A connection was meant to stand for someone a member could vouch for. “No one can have 20,000 professional connections whom they know well enough to credibly vouch for,” Hoffman wrote. After “extensive discussion and design thought,” LinkedIn capped the displayed number at 500+. The cap, in his words, “allowed us to utilize a game dynamic that tapped into pride, but also to temper it.”

1 · The plan

Greed was supposed to fill the profiles. The average ran 90 characters.

2 · What happened

Pride filled the connection counter instead. Some members passed 20,000.

3 · The fix

LinkedIn capped the display at 500+.

Figure 2. LinkedIn’s first years, as Hoffman described them in 2021. The numbers are his. The member and the profile are drawn for illustration.

LinkedIn turned down engagement in other places too. It was built on real identities, and until much later it did not let members upload photos of themselves. Photos “would have tapped into pride, lust and envy, and dramatically accelerated engagement,” Hoffman wrote. LinkedIn did not want to become “the workplace version of ‘Am I Hot or Not?’” It also chose subscriptions over advertising as its main revenue, because Hoffman wanted “a time-saving platform, not a time-wasting one.”

Sometimes a sin worked for the platform. When LinkedIn opened its publishing platform in October 2012, only 150 hand-picked figures, among them Richard Branson and Arianna Huffington, could post. LinkedIn occasionally showed them an internal leaderboard of who posted most and whose posts drew the most views. “Naturally, pride and envy kicked in,” Hoffman wrote, and the influencers wrote more. By the time every member could post, 17 months later, the early writers had set the norms.

The 2011 aphorism leaves this part out. A designer chooses which sin to feed and how hard. That choice decides whether the desire that brought someone in turns into the behavior the product was built for. The heuristic runs short in four places.

A sin fits any story

Someone posts a promotion on LinkedIn. Pride could explain the post. Greed could too, if the poster hopes a recruiter will notice. So could envy of a former classmate’s promotion last month. Or the poster might simply want to thank a team in public.

A framework that explains every behavior after the fact predicts little in advance. Hoffman’s own record shows the problem. His heuristic picked greed for LinkedIn, and greed predicted neither the 90-character profiles nor the 20,000-connection collections. Both made sense once they had happened.

The sins make a memorable list. Nobody tested them as a theory of motivation. For a founder, a sin is a hypothesis about demand. It becomes evidence when people keep using the product and pay for it, for the reason the founder proposed.

The motives the sins leave out

A scientist can spend years on a question with no commercial payoff in sight. A musician practises scales because getting better feels good. Calling either of them proud explains away the reason each would give.

Edward Deci and Richard Ryan offered another list in a 2000 paper on self-determination theory.

Self-determination theory

Deci and Ryan proposed three basic psychological needs.

Autonomy is the sense that your actions follow from your own choices and values.
Competence is the sense of being capable, especially while learning something hard.
Relatedness is the sense of real connection to other people.

A sin describes a desire in the moment. A need describes what leaves a person better off over months. Dorian Peters and Rafael Calvo review how designers have applied the theory in fields from games to health.

A language app can feed pride with streaks and badges. The need it serves is competence, and the test is whether the learner can order dinner in Spanish a year later. An accessibility tool sells convenience. Its larger effect is on autonomy, the ability to get through a day without asking for help.

Wanting a product is cheap

A hypothetical startup builds an AI app that helps people earn more. Greed is on its side and the market looks huge. Neither fact says whether anyone will open the app twice. The advice might be unreliable. The gains might be too slow to notice. A customer with a working spreadsheet may see no reason to switch, and finding each new user may cost more than that user pays.

LinkedIn’s profiles showed the same gap. Hoffman read the motive correctly, since people joined to advance their careers. He misjudged how much work they would do on the way.

Interest in a product is easy to collect. A customer who says “I’d use that” over coffee has spent nothing. Repeat use and payment are stronger evidence. The strongest is a measured improvement in whatever the customer came for.

Engagement can grow while users gain nothing

The fourth gap matters most. A product can hold attention for hours and leave its users no better off.

William Brady and colleagues measured one version of this on Twitter, in a 2021 study in Science Advances. Two preregistered observational studies covered 7,331 users and 12.7 million tweets, and two behavioral experiments added 240 participants. “Positive social feedback for outrage expressions increases the likelihood of future outrage expressions,” the authors found. Users also matched their outrage to the norms of the networks around them.

Figure 3. The loop Brady and colleagues found on Twitter. Positive feedback on an outraged post made the next outraged post more likely, across 12.7 million tweets from 7,331 users and in two experiments with 240 participants. Users also matched their outrage to their network’s norms, which the loop does not show.

Nobody at Twitter had to design for anger. A like button rewards whatever draws reactions, and outrage draws reactions.

What the outrage study measured

The study covers one platform. Likes and retweets worked like rewards in reinforcement learning, where feedback on one action shapes the next. In ideologically extreme networks, where outrage was already common, users responded less to feedback and more to the norms around them. The study measured how much outrage people expressed. It did not measure whether expressing it harmed them.

The trouble starts when a company’s metrics drift from what users came for. A social app may count longer sessions as success. Its users may have joined to keep up with friends, and an extra hour of scrolling does not show that they did.

Compare two hypothetical networking apps. One keeps a user browsing updates for an hour. The other finds the one contact worth meeting in ten minutes and lets the user go. The first wins every engagement chart. The second did the job.

When revenue rises with time spent, a company has a reason to keep users longer even when they would be better off leaving. Hoffman gives this as a reason for LinkedIn’s subscriptions. Without advertising, he wrote, LinkedIn “never felt compelled to maximize time spent on site.” Builders who fail to check “lowest-common-denominator impulses,” he argued, end up “passively optimizing for” the sins.

The best AI product may be one you rarely open

Most of the consumer internet was built around screens that people tap through themselves. An AI agent can do the tapping. It might book a trip or clear a month of paperwork while its user does something else.

Such a product gets more valuable as people spend less time in it. A hypothetical assistant that saves someone five hours of administrative work a week might be opened twice, to approve a decision. Every engagement metric would score it as a failure.

A feed

Hours of attention, and little to show for them.

An agent

Opened twice a week, with five hours handed back.

Figure 4. Two hypothetical products. Blue bars show time spent in the app, which engagement metrics count. Green bars show the hours a user gets back, which those metrics miss. The bars are illustrative.

Hoffman’s sins still explain why someone would try such an assistant. Sloth wants the paperwork gone, and greed wants the hours back to earn with. The value lies in what the hours buy. A parent may spend them with a child. A scientist may spend them on an experiment.

The same capability raises a harder problem. A system that books trips well will soon be asked which job offer to take. People who hand over every such decision may get better results and lose practice at choosing which results they want.

In Superagency, published in 2025, Hoffman and Greg Beato argue that AI can widen human agency by putting expertise within reach of more people. Design decides whether a given system does. An assistant that lays out the options and their trade-offs leaves the choice with its user. One that hides its reasoning and steers every decision can wear down autonomy even when each recommendation is accurate.

The question for agent builders

An agent succeeds outside the app, in hours returned and in decisions its user still understands. A usage dashboard records neither. What an agent company should count in their place is still an open question.

Where desire and need line up

Hoffman’s heuristic still earns its place. A technically brilliant product that pulls on no human desire usually finds no users. The sins name what pulls people in. Self-determination theory names what leaves them better off. Often the two line up.

Figure 5. How a desire can line up with a need, or fail to. Green arrows show the sin that brings a user in feeding one of the three needs in self-determination theory. Red arrows show a sin fed with nothing behind it. The pairings are illustrative and were not measured.

Recognition for real achievement can build competence. Money can buy independence. A service that removes busywork can hand back time for the people in someone’s life. In each case the desire that brought the user in and the need the product serves push the same way.

Problems start when they come apart. A product can feed pride with no achievement behind it, or envy with no way for users to improve their own position. It can stir outrage that leads nowhere. Such products can make money. Their revenue says little about what they do to the people using them.

I work on evaluating AI systems in science and medicine, where the gap is easy to see. A physician adopting a clinical AI tool wants help that can be trusted and more time with patients. Minutes spent inside the tool count against it. Its value shows up in the care, outside the product.

Hoffman’s 2011 line is memorable because it is right about why people sign up. His 2021 essay is more useful, because it records what happened next at a company he ran and what he changed. The sins answer what makes people want a product. As AI moves into decisions that matter more than what to watch next, the question that counts is what happens to people because they use it.

Understanding human desires may help us build successful products. Understanding how to serve human needs is what gives those products a reason to exist.