Trusted AI at Autodesk

Our AI transparency cards provide details on functionality, data sources, and the privacy and security safeguards in place for the artificial intelligence features used in our products.

The name of the Autodesk product and the name of the AI feature are presented at the top of the card, below the verbiage “AI Transparency Card.”

• Autodesk product name (e.g., Autodesk Forma)
• AI feature in the product (e.g., Embodied carbon analysis)

The card description summarizes the actions the AI feature is expected to perform when used within the product.

Feature functionality describes the AI and/or machine learning (ML) technology capabilities that enhance the AI feature using one of the following three terms:

• Automate: Autodesk AI reduces repetitive tasks by automating steps that have traditionally required manual work or significant overhead, minimizing error and freeing up more time for creative work and innovation.
• Analyze: Autodesk AI provides actionable insights to end users when faced with overwhelming amounts of complex data, helping them understand what is most important in real time to create the smartest solutions.
• Augment: Autodesk AI augments creative exploration and problem-solving by improving speed, quality, and breadth of thinking through contextual understanding.  

The model source describes the source type from which the model was developed to power the AI feature:

• Proprietary: The AI/ML model was developed internally by Autodesk.
• Open source: Autodesk uses AI/ML model that was developed by a third party, who made it available to the public.
• Licensed: Autodesk has a license to use the AI/ML model that was developed by a third party.
• Combination: Part of the AI/ML model was developed internally by Autodesk, and the other part(s) were developed by a third party (open source and/or licensed).

The models behind each AI feature use methods, approaches, and techniques to learn from data, find patterns, perform tasks, and produce outcomes. We use techniques that will improve the quality and value of our products for customers. Techniques are constantly evolving, and in some cases multiple techniques are used, some of which may not be listed here. This field describes the primary technique used to develop the AI feature:

• Transformer: A machine learning technique designed to process and understand data to perform sequential tasks more efficiently, such as language translation.
• Encoding: A process of converting data into a specific format that can be efficiently processed by machine learning models.
• Classification: A supervised learning technique that assigns items into predefined categories and predicts the category of new observations based on historical data.
• Feed forward neural network (NN): A deep learning technique where information flows in one direction, from input to output, without any cycles or loops.
• Predictor: AI technique that learns from data to make informed predictions about future events or outcomes based on historical data and patterns, such as forecast results, make decisions, and provide insights.
• Genetic Algorithm: a method for solving both constrained and unconstrained optimization problems that is based on natural selection concepts.
  • Constrained optimization problems use logical limits or conditions that a solution to a problem must consider. They reflect real-world limits on things like production capacity, inventory, available space, and so on.
  • Unconstrained optimization problems deal with situations where there are no predefined limits or conditions for a solution to consider.
• Transformer diffusion: A transformer technique (see Transformer above) that creates data by reversing a diffusion process by gradually adding noise to the data.

Indicated with a "yes" or "no" designation, this describes whether the generated output can be reviewed and/or further updated by the user before any final action is being taken. This is otherwise known as "human in the loop."

This section indicates whether personal data is present in the training, testing, or validation datasets used for the development of this feature. 

The data source designation listed in the card identifies the types of data sources used for the development of this feature. This includes the data that was used to train the model that powers the AI feature. The types of sources are categorized as follows:

• Open source:  Data that is freely available for use, modification, and distribution under an open license.
• Customer content: Data that the customer or their authorized users submit or upload to the product, as further defined in Autodesk’s Terms of Use as “Your Content.”
• Synthetic data: Data generated by a system or model that can mimic and resemble the structure and statistical properties of real data.
• Commercial: Data that is purchased and/or acquired from a third party under a restrictive license.
• Mix: More than one data source category was used.
• Customer trained: The customer performed training and used their own proprietary data.

Choice formats are indicated as Opt-in/Opt-out, No, or N/A. These labels identify the form of choice available to customers and/or their users when their data is used for the AI feature's development/improvement.

• Opt-in/Opt-out: The customer can choose whether to opt in or out of the use of their data for feature development/improvement.
• No: A choice is not offered.
• N/A: A choice is not applicable, as no customer content is used for feature development/improvement.

We provide information about two types of encryption: encryption at rest, and encryption in transit. Both of these are shown with Yes or No designations.

• Encryption at rest: Indicates whether the data is encrypted in the database(s) where the data is maintained. All encryption at rest uses the Advanced Encryption Standard (AES) 256-bit key length, otherwise known as AES-256.
• Encryption in transit: Indicates whether the data is encrypted as it's transmitted from one point to another. Autodesk enforces encryption in transit via HTTPS standard encryption, RSA with AES-256, using TLS 1.2 at minimum.

This section of the card indicates, as applicable, what other notable mechanisms are employed to preserve the confidentiality and protection of the data in addition to our standard security mechanisms. These safeguards apply to both personal data and company data.

• Tokenization: Sequences of information in data are broken down into smaller units called “tokens.”
• De-identification: Identifiers are removed from data and replaced with placeholder values.
• Anonymization: The dataset does not contain any identifiable information and there is no way to link the information back to identifiable information.