liveBook · Manning

Alignment training

As explained in Causal AI

AI alignment is the task of making an artificial intelligence system's behavior consistent with human goals and preferences.

Alignment training makes an AI system’s behavior consistent with human goals and preferences. In causal analysis, that means moving beyond objective causal facts toward judgments people consider appropriate, such as responsibility. section 13.1.4

The distinction matters when objective causality is straightforward but interpretation is subjective. An LLM may identify events associated with a network slowdown, then apply human-specified criteria, legal theories, or case law to produce a judgment. section 13.1.4

Interpreting a network slowdown

A log shows Jake logging in, accessing a shared folder, launching a heavy process, and coinciding with a bandwidth spike and packet loss. Alignment asks the LLM to assess Jake’s responsibility in ways many humans would accept, rather than merely confirming the causal sequence. section 13.1.4

Aug 22 14:02:34 User "Jake" logged in from 192.168.0.105.
Aug 22 14:03:12 Shared folder accessed by user "Jake" at //SERVER/shared-folder.
Aug 22 14:03:20 File transfer initiated by "Jake" from local: /home/jake/documents to remote: //SERVER/shared-folder/documents.
Aug 22 14:03:25 WARNING: Unusual network activity detected. User "Jake" launched file "heavy_process.exe" on //SERVER/shared-folder.
Aug 22 14:03:30 Network performance alert: Bandwidth usage spikes at 95%.
Aug 22 14:04:00 Network slowdown detected: Packet loss rate 35%.

Inspect the event sequence used to connect causal evidence with a human-aligned responsibility judgment. section 13.1.4

How Causal AI covers it

How deep it goes
Defines alignment through causal responsibility and instructed criteria, emphasizing judgment rather than parameter-level training procedures.
Examples
Uses network logs, a bandwidth spike, packet loss, and an LLM judgment about Jake’s responsibility.
What it assumes
Assumes familiarity with causal facts and LLMs while distinguishing objective causality from subjective responsibility.
What you can do after
Apply it via prompting by specifying criteria, legal theories, or case law for an LLM’s judgment.
How it connects
Connects large language models with causal inference, causal judgment, root-cause analysis, and legal reasoning.
Cover of Causal AI
liveBookmanning.com
Read Causal AI on liveBook Get the book at manning.com

Also explained in

Compare how 6 books explain alignment training →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Causal AI. The section links point to the parts of the book it draws on.