I refuse to use “AI”, and you should boycott it too. It is an evil technology entirely based on theft, with massive ethical, legal, societal, environmental, and financial problems. This thread explains why. 🧵

The training data for these language models is entirely based on stolen data. Art, books, code, knowledge – all stolen from their authors without their knowledge or consent. It is the largest theft of intellectual property in history.

‘Doom Loop’: OpenAI and Microsoft Admits LLMs Are Destroying the Web and Built on Theft

“Millions of people around the world will soon consider large models ‘hoovering up’ all their work to be an astonishing theft of unprecedented proportions.”

Unlike a human training by studying existing media, AI models simply copy vast swatches of copyrighted material into themselves, and can even be tuned to regurgitate those works verbatim.

Nav Toor (@heynavtoor) on X

🚨BREAKING: Every book you have ever read. Every novel that has ever been published. It is sitting inside ChatGPT right now.

Word for word. Up to 90% of it. And OpenAI told a judge that was impossible.

Researchers at Stony Brook University and Columbia Law School just proved it.

They fine tuned…

Major AI companies have trained their models by feeding them vast swathes of pirated books. While they were eventually taken to court over this criminal act, they got off with a mere payout, while keeping the models trained on this stolen data.

AI Firm To Pay $1.5 Billion For Pirated Books, Keeps Its Trained Models

Anthropic to pay $1.5 billion to settle a major copyright case over pirated training data—but the company gets to keep its Claude models intact. No deletion. No rollback.

They then moved to training models using old physical books, with AI companies buying second-hand books en mass (including very rare books), scanning them, and then destroying the originals.

Rare Books: Why AI Scans and Destroys Them

AI companies buy rare books in bulk, scan them with machines that cut the spine, then destroy the original: a judge ruled the practice legal as fair use.

Models for image generation were trained on hundreds of millions of stolen images which were scraped from the internet, again without the artists’ knowledge or consent.

Midjourney Founder Admits to Using a ‘Hundred Million’ Images Without Consent

It has outraged artists and photographers.

Models and agents for programming are also based on plagiarism, with training largely taken from public source code repositories, including ones under copyright (or copyleft.) Again, code authors never consented to this usage, and many repos predate LLMs.

GitHub’s Commercial AI Tool Was Built From Open Source Code

Copilot is pitched as a helpful aid to developers. But some programmers object to the blind copying of blocks of code used to train the algorithm.

More general models, such as those used by chat agents, were largely trained on data stolen from forums, social media, articles, and wikis – once again without the authors’ consent.

Every Prompt You’ve Ever Typed May Be Training an AI Model — Without Your Consent

In 2020, OpenAI released GPT-3. To train it, they used a dataset called The Pile — a massive corpus…

While I’ve singled out specific examples in this thread, ALL major AI models (including OpenAI / ChatGTP, Anthropic / Claude, Microsoft / CoPilot, and Amazon, to name a few) have been trained in this way, on human-made data taken without the permission or consent of the authors.

If you are a creative and you are considering using AI, remember that it is a technology that is designed around stealing the work of other people just like you.

There are also a lot of unresolved legal issues around the use of generative AI. Anything made using it may be at risk of legal action in the future, once the laws finally catch up with the technology.

The main legal defence from AI companies is the claim that training AI models on copyrighted material counts as “fair use”, which is not proven law at all, with some courts already ruling that it is not (at least in certain circumstances.)

Court Finds That Using Copyrighted Material as AI Training Data Is Not Fair Use

A US appeals court upheld a ruling that AI legal research company Ross Intelligence’s training of its AI model didn’t amount to “fair use.”

As already mentioned, AI executives have admitted that the technology is based on theft.

They are spending vast sums of money to lobby lawmakers into making exemptions for them, especially in the USA. It remains to be seen if they will be successful.

The AI Lobbying Boom 2026: How Tech Giants Are Buying AI Regulation | OpenLobby | OpenLobby

AI lobbying spending surged 340% since 2023. In Q1-Q2 2026, Google, Meta, OpenAI, Anthropic, and Microsoft have spent record amounts to shape AI regulation before Congress acts.

The environmental costs of the data centres powering AI are staggering. The power used by data centres, the water used to cool them, and the resources used to build them are massive. Then there is the manufacturing of all the hardware they need.

A UN report estimates that AI data centres will consume more power than 650 million people, and more water than 1.3 billion people, by 2030.

Rising Emissions, Depleting Water and Vanishing Land—UN Scientists: AI Is Threatening Natural Resources for Billions

By 2030, AI’s water use will match the needs of 1.3 billion people while its power use triples that of 650 million, UN University investigation warns

Data centres are expected to account for the largest increase in energy demand by 2030, especially in the USA – all at a time when we urgently need to reduce our dependence of fossil fuels to avoid climate catastrophe.

Electricity demand growth by sector and end-use in the United States, 2015-2030 – IEA

Electricity demand growth by sector and end-use in the United States, 2015-2030 – Chart and data by the International Energy Agency.

Overall, fossil fuels (gas and coal) power over half of data centres in both the USA and China. While this is expected to shift more towards renewables in the future, the current policies in the USA greatly favour the more polluting sources of power.

Energy supply for AI – Energy and AI – IEA

Energy and AI – Analysis and key findings. A report by the International Energy Agency.

In the USA, there has been talk of building new gas pipelines specifically to power AI data centres.

Gas pipeline operators in talks to fuel AI datacenter demand

Utility tapped out? Why not build your own?

In addition to the environmental effects, the energy demand of data centres has also increased the cost of energy for everyone. In the USA, data centres are estimated to have already cost the public $23 billion in price increases.

Data centers have already hiked electricity prices on the public by $23 billion. Good luck clawing that back | Fortune

A PJM market watchdog calls the shift a “massive wealth transfer” to tech companies, but there’s a peak-demand loophole.

Here in Ireland, citizens are estimated to have paid an extra €715 million (approx. €300 per household) on electricity between 2015-2023 due to data centres, with that amount expected to more than double over the next decade.

The cost of data centre growth in Ireland: Households paid an estimated €715 million more in electricity bills

Ireland’s current approach to data centres is increasingly a cost-of-living issue

The demand for hardware for AI data centres has also led to unprecedented price increases for computer hardware, something which gamers and game developers have felt especially hard as they need high-performance PCs.

Hardware Price Increases: AI Is Quietly Taking Hardware Away from Everyday Users

Hardware price increases at a fast speen in this year. SSDs, HDDs, RAM, GPUs, and even USB drives are getting more expensive. Learn why AI is reshaping hardware supply, and what it means for you.

As an aside, hardware manufacturers such as Nvidia, Micron, and AMD are the only companies actually making a profit from AI, because they are the ones selling the hardware to everyone else.

Source: https://isaiprofitable.com/

The major AI companies have been created and run by some of the most evil human beings on the planet. Apart from the grand-scale theft that’s already been covered, they have a laundry list of other offences (some confirmed, some merely alleged.)

After former OpenAI (creators of ChatGPT) employee Suchir Balaji blew the whistle on the mass theft involved in training, many believe that OpenAI either hounded him to suicide, or had him murdered to silence him.

OpenAI whistleblower death: Autopsy suggests murder, not suicide

Former OpenAI employee Suchir Balaji made headlines for his revelations about ethical concerns surrounding generative AI

In 2023, OpenAI’s head of research, Tal Broda, openly called for the genocide of the Palestinian people. While he later deleted the posts, he never suffered any repercussions from OpenAI.

Tal Broda – Complicit in Apartheid | Reverse Canary Mission

Tal Broda, a senior OpenAI executive, weaponised his platform to echo calls for the continued ethnic cleansing of Palestinians through deleted tweets that …

OpenAI’s founder, Sam Altman, has been accused of sexually abusing his sister while she was a child.

OpenAI boss Sam Altman accused of sexual assault by sister in lawsuit

Altman has denied the allegations in a statement, labelling them “utterly untrue”.

Elon Musk, founder of XAI, performed a Nazi salute twice during Donald Trump’s inauguration rally.

Ruth Ben-Ghiat (@ruthbenghiat.bsky.social)

Historian of fascism here. It was a Nazi salute and a very belligerent one too.

[contains quote post or other embedded content]

Major AI companies OpenAI, Microsoft, Google, XAI, Nvidia, Microsoft, Amazon, ReflectionAI, all signed a deal with the Pentagon allowing use of their AIs for all purposes, including fully autonomous weapons and the mass-scale surveillance of citizens.

Pentagon signs classified AI deals with Nvidia, Microsoft, and AWS after ejecting Anthropic over safety limits

Seven companies now operate AI on classified Pentagon networks under “lawful operational use” terms. Anthropic, which refused to drop its safety limits, was replaced.

Amazon, Google, Microsoft, and possibly other American AI companies, worked with Israel in building AI systems to create kill-lists of Palestinian civilians to murder during their genocide of Gaza.

Israel’s AI use in Gaza potentially normalizes civilian killings, obscures blame, exposes Big Tech complicity: Expert

Israel is using AI systems with known inaccuracy risks at ‘almost every stage’ of its military operations, says Heidy Khlaaf, chief AI scientist at AI Now Institute | Anadolu

Anthropic’s Clause AI was used by the USA and Israel when conducting their unprovoked bombing of civilian targets in Iran.

Anthropic AI reportedly used in US-Israel strikes as OpenAI faces user backlash for military agreement – Business and Human Rights Centre

Check out this page via the Business and Human Rights Centre

I’ve released my second Unreal Engine Marketplace product.
This one is a configurable randomisation system, written entirely in Blueprints.

It allows random sequences of either numbers (e.g. die rolls) or outcomes (e.g. chance to hit) to be generated, with options to coax the results into fairer distributions and to limit unlikely edge cases.

We’ve all played games where some randomised result feels really unfair because it seems to keep rolling against us repeatedly. Since the outcome of each roll is not affected by previous rolls, it’s perfectly possible (though unlikely) to get any number of low-probability results in a row, and given enough attempts and enough players, such a sequence is almost guaranteed to happen to someone.

So, I created this system to allow designers to control the randomness, to make it slightly less random but also closer to what we as humans expect random outcomes to be like, based on some of the tricks I’ve read about other game developers using.

It uses a two main methods to achieve this effect:

The first method is the use of a shuffle bag – instead of rolling for each outcome as needed, a number of future outcomes are placed into a bag and their order is randomized.
For example, given a 20% chance, two successes and eight failures are placed into the bag, guaranteeing an exact 20% outcome over ten rolls.
This actually changes the nature of the what is being randomised – the individual outcomes are no longer random, but instead the order of the outcomes is randomised.

The second method uses random generation, but adjusts the chances to favour some outcomes and disfavour or disallow others (such as excessively long streaks.)
For example, when rolling six-sided dice, each number can become more likely to be rolled the more rolls it has gone without appearing.

The system has a lot of options to further configure the randomisation algorithm. The full documentation is available on Google docs.

These tricks can be used for things like making sure the player doesn’t go too long without scoring a hit or a critical strike, to make sure that a weapon which deals damage in a range has an even spread inside that range, or to make it increasingly unlikely for a player to go without finding rare treasure with a very low chance to drop, to name but a few examples.

NV Random on the Unreal Marketplace

This tutorial is a quick summary of the process for preparing a content pack for submission into the Unreal Marketplace.
It is geared towards content-based asset packs.

The tutorial will cover three main aspects:

1. Preparing the content, including a list of things to check before submission.
2. Submitting the product to the Marketplace.
3. Updating a submission, including updates when a new version of the Unreal Engine is released.

Read the rest of this entry »

The goal of this tutorial is to set up a lantern object which casts soft dynamic shadows from the solid parts of the lantern.

The two normal ways of creating detailed shadows are to either bake them with static lights (which will not work for a moveable light source), or to use Raytraced Distance Field Soft Shadows, which don’t work very well if the light source is actually inside the model which is supposed to cast the shadow.

Instead of either of those, I will explain how to solve this problem by using a light function.

A light function is a material shader that can filter the intensity of light being output by a light source.  It can be used to give the light different intensities at different angles, with the output of the light dependant on the colour of the material, from black (no illumination) to white (full illumination).

We will use this to effectively mask the parts of the light where we don’t want it to shine at full brightness.
Further, we will blend between two different masks depending on the distance, giving a more realistic effect where the shadows are fairly sharp close to the light source, but get softer as the distance increases.

Read the rest of this entry »

The following are some musings on how to make good RPG systems, and which mistakes or pitfalls to avoid, based on the systems I’ve seen in various roleplaying games.

Note that I am talking specifically about RPG character progression systems, and not the other staples of the genre such as dialogue, choices and consequences.

I feel that a character system should:

• Have interesting and meaningful choices
• Differentiate gameplay between character builds
• Feel satisfying
• Avoid offering truly bad choices which only act as traps for inexperienced players
• Not get in the way of normal play

Read the rest of this entry »

I have played a lot of games, and in so doing I have seen a lot of the same mistakes, or at least design decisions that I disagree with, repeated many times.
I have also seen many excellent ideas in games that really helped them to stand out and be exceptional.

Recently, I’ve taken some time to think about the types of things that I do and don’t like to see in games, and from there I have created a list of what I consider to be good game design goals.

Of course, some of these are subjective – they are for making the type of games I like to see – but most of them are fairly general and should apply to almost every game.

Game Design Goals

1. The player should have to think.

While some amount of direction is necessary, the player shouldn’t be spoon-fed exactly what to do next every second, or every time that they complete an objective.  Occasional hints are fine, but the player should need to think and work out how to achieve their goals for themselves, rather than just following instructions.

Bad: The player is told exactly what to do each step of the way.
Good: The player is told roughly what to do, and can find information hinting at the solution, or work it out for themselves.

Read the rest of this entry »

When creating levels for games, you will often want to place sequences of terrain features which repeat, for example: stairs, fences, and other repetitive details.

When making stairs, you could model an entire staircase in your 3D-modelling package specifically for the area at hand, but that requires you (or your artists) to make a lot of customised, single-use meshes.

Alternatively, you could model individual steps (or a mesh with a small number of steps) and repeat those until your staircase is the right height.  Constantly using the same model will most likely make the repetitive use of the same texture extremely obvious and jarring.

This article proposes a solution to this problem: to create a number of unique meshes which are all variations on the same section of the repeating terrain, and to then use a blueprint to automatically select and place the individual meshes to build up the desired length of terrain.

For example, to model a dozen variations of a single step, and then have a blueprint which randomly (or sequentially) picks one of those variations and places it appropriately until it has built the entire staircase.

This allows the level designer to quickly and easily create repetitive terrain features, each with a unique look, and without requiring custom assets to be made for each individual use case.

Four sets of stairs generated from four step models using the system described in this tutorial.

Read the rest of this entry »

Bioshock is a 2007 game created by Irrational Studios, and is considered to be a spiritual successor to System Shock 2 (1999).

Despite its unique setting and atmosphere, the game contains a number of major design flaws that are worth investigating as game design anti-patterns (anti-patterns are the study of a “wrong” solution to a problem, to understand why it’s wrong and how the problem could be solved in a better way.)

In this article, I am going to specifically talk about two areas of Bioshock’s design;

• Interface – the interfaces that the game uses to allow the player to control its various aspects, such as genetic powers and items.
• Balance – the way that the various elements of the game are balanced against each other, and the way the game provides difficulty to the player.

Before I begin, I would like to mention a “pitch document” from early in Bioshock’s development, which set out the developers’ aims and design goals for the game.
This document is available as a set of images on Irrational’s Website.  I will be referring to and quoting sections of this document throughout the article, as I describe the various problems that the developers were looking to solve.

Read the rest of this entry »